artificial intelligence in power station research paper

• Search Menu
• Sign in through your institution
• Advanced Articles
• Editor's Choice
• Author Guidelines
• Publish with us
• Submission Site
• Open Access
• Self-Archiving Policy
• About Clean Energy
• About the National Institute of Clean and Low-Carbon Energy
• Editorial Board
• Instructions for Reviewers
• Advertising & Corporate Services
• Journals Career Network
• Journals on Oxford Academic
• Books on Oxford Academic

Article Contents

Introduction, 1 methods and challenges addressed, 2 ai techniques, 3 application of ai in the power sector, 4 conclusions, 5 future scope, authors’ contributions, conflict of interest statement, data availability.

• < Previous

Applications of artificial intelligence in power system operation, control and planning: a review

• Article contents
• Figures & tables
• Supplementary Data

Utkarsh Pandey, Anshumaan Pathak, Adesh Kumar, Surajit Mondal, Applications of artificial intelligence in power system operation, control and planning: a review, Clean Energy , Volume 7, Issue 6, December 2023, Pages 1199–1218, https://doi.org/10.1093/ce/zkad061

• Permissions Icon Permissions

As different artificial intelligence (AI) techniques continue to evolve, power systems are undergoing significant technological changes with the primary goal of reducing computational time, decreasing utility and consumer costs and ensuring the reliable operation of an electrical power system. AI techniques compute large amounts of data at a faster speed than numerical optimization methods with higher processing speeds. With these features, AI techniques can further automate and increase the performance of power systems. This paper presents a comprehensive overview of diverse AI techniques that can be applied in power system operation, control and planning, aiming to facilitate their various applications. We explained how AI can be used to resolve system frequency changes, maintain the voltage profile to minimize transmission losses, reduce the fault rate and minimize reactive current in distributed systems to increase the power factor and improve the voltage profile.

The demand for advanced research and technology has steadily increased in the sector of electric grids [ 1 ]. Automation and intelligent technology have become widely used in response to development demands over time. Traditional research methods are quickly becoming insufficient to enable data scientists and researchers to keep up with any global challenges that artificial intelligence (AI) may be able to assist us in solving and uncovering important insights among the billions of pieces of data scattered throughout power systems. AI can handle large amounts of data and utilize them to make power system operations, control and planning more efficient. The use of AI technology in power systems has been investigated and debated in related areas and has resulted in more study material and certain outcomes, which are reviewed in this paper. The demand for advanced research and technology has constantly risen in the sector of electricity grids. The application of AI technology to the automation of power system control can improve the efficiency of electrical automation management, mitigate the risk of accidents and ensure smooth operation of the power system over an extended period [ 2 ]. Evaluating the use of AI technology in power systems requires a comprehensive analysis of existing research in the field of artificial intelligence and its related industries.

The power system is a network consisting of three components: generation, distribution and transmission. In the power system, energy sources (such as coal, sunlight, wind, nuclear reactions and diesel) are transformed into electrical energy [ 3 ]. There are different power systems, such as solar power systems, wind power systems, thermal power plants [ 4 ], nuclear power plants, geothermal power plants, etc. All power systems have different structures and equipment for the generation of electricity [ 5 ]. The basic structure of a power system includes:

(i) generating substation;

(ii) transmission substation;

(iii) sub-transmission substation;

(iv) distribution substation.

In power system problem-solving, conventional approaches such as practical numerical optimization methods (e.g. lambda iteration and Newton–Rapson methods) have been used. Optimization problems are non-linear and, with the various constraints included, these optimization problems become slow and complex. So, several AI techniques are discussed here to solve many optimization issues with less computation time. Furthermore, experiments were conducted to identify which backpropagation algorithm would give the most efficient and reliable network training [ 6 ]. The systematic approach was introduced [ 7 ] for developing a recurrent neural network (RNN) that could effectively predict the dynamic behaviour of a pilot-scale entrained-flow gasifier. The RNN was trained using a data set consisting of input and output data collected from a dynamic read-only memory (ROM) system established in a previous study. To evaluate the accuracy of the RNN, comparisons were made with computational fluid dynamics models and experimental data obtained from the pilot-scale gasifier. The findings demonstrated that the RNN surpassed the performance of the dynamic ROM model, showing strong predictive capabilities to capture the transient behaviour of the gasifier. Additionally, this was achieved while significantly reducing the computational resources required. Power systems are one of the main study topics for the advanced development of AI. The use of AI algorithms in power plants has been under continual investigation since the advent of the expert system technique [ 8 ]. However, in most situations, issues such as long cycle times, complex computation and difficulty in learning arise with classic AI methods. In recent years, efficiency has increased significantly with the continued advancement of AI algorithms. The multi-source model of heterogeneous large data has gradually developed, the data volume continues to grow, and new possibilities and problems are being created with the use of AI in power systems. AI encompasses several technologies such as expert systems, pattern recognition, genetic algorithms (GAs) and neural networks. By incorporating AI into the automation of power system control, it has the potential to enhance the efficiency of electrical automation management, mitigate the risk of accidents and ensure long-term smooth operation of the power system. Machine learning (ML) has also found extensive applications in predicting the properties of rechargeable battery materials, particularly electrolyte and electrode materials, as well as the development of novel materials, according to relevant research. The breadth of ML applications will grow steadily as ML technology advances and new unique issues emerge in the research of rechargeable battery material [ 9 ]. Although ML has shown considerable potential in modelling complex systems, its implementation introduces new challenges. These challenges include difficulties associated with accessing relevant and reliable data sets, and addressing inaccuracies in model predictions needs to be addressed before ML can be widely deployed. To effectively implement ML on a large scale, it is crucial to have compelling evidence of its effectiveness in diverse areas such as manufacturing processes, energy generation, storage and management. Furthermore, the availability of commercial software and a skilled workforce specialized in the relevant domains is essential [ 10 ]. ML has proven to be beneficial in creating data-driven models that accurately correlate material properties with catalytic performance, including activity, selectivity and stability. As a result, there have been advancements in the development of effective design and screening criteria for solid-state catalysts with desired properties [ 11 ]. Nevertheless, there are still difficulties in applying existing ML algorithms to accurately predict catalyst performance or devise strategies for designing high-performance catalysts. The review emphasizes recent advancements in ML applied to solid heterogeneous catalysis, as well as the limitations and constraints faced by ML in this field, and also discusses some of the prospects for using ML effectively in the design of solid heterogeneous catalysts. Successful uses of ML in short-term hydrothermal scheduling will strengthen the link between real operations and issue formulation, and prepare the hydropower sector for autonomy by identifying the need for and availability of autonomous systems now and in the future. In this research, a review of the state of the art of ML applications for the hydroelectric sector was offered [ 12 ]. Digital technologies have a significant impact on energy market services and the safety of residents and energy consumers, particularly in smart homes. Sustainable smart home networks can improve energy efficiency, utilize local renewable energy, decarbonize heating and cooling systems, and promote responsible electric vehicle charging [ 13 ]. The next decade is crucial for achieving ambitious global CO 2 reduction targets and the decarbonization of buildings is a major challenge. Water-efficient development and resilient homes are essential for coping with impacts of climate change. Research on sustainability and energy efficiency is vital to improve the quality of life in the face of climate change [ 14 ]. Several important aspects were highlighted [ 15 ] with respect to the current state and prospects of smart homes. It was recognized that despite the increasing prevalence of smart homes and the growing familiarity with them, there were still significant obstacles that researchers must address to achieve widespread adoption [ 16 ].

One of the technical hurdles highlighted is the diversity of manufacturers and devices, each with varying charging systems, frequencies and communication methods [ 17 ]. This fragmentation can hinder interoperability and compatibility between devices and systems. The greater acceptance of smart home technologies poses a significant challenge. This review highlights the crucial task of convincing consumers about the safety and reliability of these technologies. An approach [ 18 ] was presented to develop prediction models that were capable of identifying faults and malfunctions in power equipment, demonstrating their effectiveness in predicting the progression of degradation phenomena. The challenges were discussed [ 19 ] associated with predicting the technical condition index of the equipment and determining the probability of its current state having defects. This research contributes significantly to the advancement of predictive analytics tools in the industry, enabling proactive maintenance of equipment. ML and data-driven approaches exhibit significant promise in the field of predictive analysis within power systems, especially in the context of smart grids. These methods can efficiently analyse the vast amounts of data collected from smart meters and other devices in real time, facilitating optimized energy flow in an increasingly renewable-energy-focused landscape [ 20 ]. They offer advantages such as improved accuracy, cost reduction and improved efficiency. However, certain challenges must be overcome, such as ensuring the availability of high-quality data and managing the potential risk of information overload [ 21 ].

The articles selected for review are based on different parameters and selection criteria. The shortlist is based on parameters such as duration, analysis, comparison and applications, as listed in Table 1 . The challenges investigated for power system operations, control and planning in the article are as follows and a diagram visualizing the domains of the power sector along with the AI techniques used and their application is presented as Fig. 1 . Power system operation [ 22 ] includes the total power requirement that must reliably meet the real-time generation, including transmission losses. The problems involved in this task are economic load dispatch (ELD), power flow, unit commitment and generator maintenance schedule.

Selection criteria for shortlisted research papers

Visual depiction of power sector domains, their application and AI techniques used

The complex and large design of the power system is presented [ 23 ] and interferences in the power system are a problem. When a large interference occurs, control tasks are needed to find the disturbed area, control the impact caused and bring the process to normal form. Heuristic solutions are non-linear and hence are not designed to deal with fast-occurring disturbances. Therefore, many control optimization techniques such as voltage control (VC), power system stability control and load frequency control are discussed to address this problem.

Power system planning has an arrangement of a power system that is complex and large with many parts such as flexible alternating current transmission system (FACTS) devices and distribution systems. The major goal of least-cost planning is to optimize the components required to deliver enough power at a minimal cost. Many factors such as FACTS placement and demand are given importance in the expansion of power system planning. Reactive power optimization, distribution system planning (DSP) [ 24 ] and capacitor placement are the optimization problems considered in this task [ 25 ].

2.1 Artificial neural network

In AI, a set of inputs is transformed into an output using a network of neurons. A neuron produces a single power by simply operating its input in the same way as a processor [ 26 ]. The working group of neurons and the pattern of their connections may be utilized to build computers with real-world issues in model recognition and pattern categorization. As the human brain processes are replicated, input signals are processed using mathematical operations utilizing artificial neurons.

The network consists of neurons organized in layers and connected to ensure information input–output flows [ 27 ]. By using what is known as the activation function, in layer ‘i’, each neuron is linked to the ‘i+1’ layer of all neurons. The input signals for a specific neuron originate from all neurons in the prior layer and their excitation power changes to govern the degree of signal reaching each neuron [ 28 ].

In several scientific disciplines, such as medical diagnosis, voice, pattern recognition, etc., artificial neural networks (ANNs) are utilized. The ANN is a computing system partly based on biological neural networks, expressed by linked nodes (artificial neurons), correctly structured in layers that are found in human or animal brains. All artificial neurons are linked and are able, employing their connections (synapses) to send signals, generally real values, which result in an output computed according to the original input, depending on the sizes allocated to all neurons [ 29 ].

ANNs are recognized as data-mining approaches capable of modelling several independent characteristics with dependent functions in non-linear functions. ANNs may predict a future value of a dependent variable after training with a comparable sample, replicating the learning process of a human brain [ 30 ].

In turn, a difference in signal strength affects the activation of the neuron and, as a result, signals that are transmitted to other neurons as shown in Fig. 2 .

Simplified diagram of the artificial neural network

Input layer—distribute other units but does not process the data.

Hidden layer—the ability to map the non-linear problems is provided through hidden layers.

Output later—the output units encode the value to be assigned to this instance.

2.2 Adaptive neuro-fuzzy interference system

The adaptive neuro-fuzzy interference system (ANFIS) creates an input/output data set whose membership function parameters are modified with the minimum square method type or the backpropagation algorithm by itself, using a fuzzy interference system [ 31 ]. This modification has helped the fuzzy system learn from the data it models. By applying hybrid learning, ANFIS utilizes a systematic approach to determine the optimal distribution of membership functions, enabling effective mapping of the relationship between input and output data [ 32 ]. The ANFIS architecture combines ANNs with fuzzy logic, making the modelling process more structured and less dependent on expert knowledge. This inference system is constructed using five layers in its basic form. Each ANFIS layer has several nodes defined in the layer specification using the node function. The current layer inputs from the preceding layer nodes are collected. The structure of the ANFIS is shown in Fig. 3 [ 33 ].

Simplified diagram of an artificial neuro-fuzzy interference system

The fuzzy-inference method involves organizing empirical information in a professional manner, which presents challenges in quantifying it through membership functions (MFs) and fuzzy rule bases [ 34 ]. Additionally, neural networks possess learning capabilities. From top to bottom, they are very adaptable in their system set-up and have great parallel processing and fault tolerance. The theories for neural network neuro-fuzzy systems are actively explored in several areas [ 35 ].

The utilization of a neuro-fuzzy system, which emulates human learning and decision-making abilities, can lead to varying model performances compared with traditional mathematical approaches. The process of rule generation and grouping in a neuro-fuzzy inference system, adapted to the specific model, can be approached through a grid-based methodology, known as the ANFIS.

2.3 Fuzzy logic

To identify the fuzzy set from which the value comes and the degree of membership within that set, fuzzy logic systems base their choices on input in terms of variables generated from the member functions [ 36 ]. The variables are then combined with IF–THEN language requirements (fuzzy logic rules) and a fluid implication is used to answer each rule [ 37 ]. The response of each rule is weighted according to the confidence or degree of the inputs of each rule and the central part of answers is computed to provide a suitable output and achieve the compositional rule of deference. Now, the design of fuzzy logic systems is not a systematic approach. The easiest way is to subjectively define member functions and rules with a human-operated system or an existing controller and then test the design for the right output. If the design fails the testing, the MFs and/or rules should be changed. Recent investigation directions involve the creation of fuzzy logic systems that can learn from experience.

Currently, only published findings can create and modify fuzzy control rules based on experience [ 38 ]. Among them can be Scharf’s self-organizing robotic control system [ 39 ] by using a performance matrix to change the rule matrix and alter the rules that constitute the management strategy. Another intriguing example is the Sugeno fuzzy vehicle, which can be trained to turn and park itself. Instead of a membership function, the effect of a rule is viewed as a linear equation of the process state variables. Through optimization of least-squares performance indices using a weighted linear regression system, the challenge is simplified to a parameter estimate. Although these approaches provide promising outcomes, they are subjective, somewhat heuristic and depend on trial and error for the choice of MFs. Thus, the ability to learn neural networks can offer a more promising approach to fuzzy logic systems [ 40 ]. As shown in Fig. 4 , the fuzzy logic system consists of four parts: knowledge base, fuzzification, inference and defuzzification. On the basis of the fuzzy constants provided, the process converts given inputs to the fuzzification stage. Based on the knowledge base, the inference is made. Then, in the defuzzification stage, every fuzzy output is mapped to complex output MFs [ 41 , 42 ].

Fuzzy logic system

2.4 Ant colony optimization

Ants exhibit a behaviour in which they remember and follow a specific path between their colony and a food source. They achieve this by leaving pheromone trails during their food search [ 43 ]. When other ants come across these pheromone trails, they start to follow them. Because the increased presence of the chemical on the path has the effect of attracting more ants to follow it, the ants will emphasize the pheromone trail. To find the best solution to the problem studied, ant colony optimization (ACO) builds multiple iterative solutions. The objectives of [ 44 ] were to evaluate the features of the search area for problems and to use this knowledge to address the solution process. The solution–construction process is a sequential decision-making process due to parametric stochastic decisions. An ACO algorithm depends on a sequence of learning of the parameters used in decision-making to reach a global policy that provides optimum solutions for a particular situation [ 45 ]. The parameters of the learning object are considered pheromones and are called variables of the pheromones.

An ACO method contains a stochastic local search technique to organize the routing pathways that artificial ants can determine. These ants co-operate together through indirect information exchange to create the best and shortest route. The concept of the ACO is taken from the food search characteristic of the true colony in an intelligent optimization algorithm and how the ants work together in this difficult job. It can be expected that the ACO finds the quickest route from nest to food according to the biological study of the ants. The ant pheromone distribution technique is termed staggered, in which information is shared with other ants indirectly. Pheromone updates are the basis of the ACO algorithm. These pheromone updates depend on the pheromone and the number of ants that work best. Natural ants can determine the quickest route based on their best knowledge and a strong pheromone trace. The shortest path is inversely proportional to the amount of pheromone and length of the path using an ACO method. The following is a step-by-step explanation of the algorithm replicating these properties [ 46 ]. The pseudocode for ACO is shown in Table 2 .

Pseudocode of the ant colony optimization algorithm

Set pheromone pathways: The algorithm starts by setting the initial pheromone pathways in the search space of the problem. These pathways act as a guide for ants to navigate and find solutions.

Generate a random ant population: Next, the algorithm generates a population of random solutions (ants) to start searching for the optimal solution.

Choose the optimal position: Each ant then uses a combination of pheromone information and heuristics to determine the next step (position) to take. The objective is to find the position that maximizes the target function.

Get the finest search ant: After all ants have completed their search, the algorithm selects the best ant, i.e. the one with the highest value of the target function.

Restore the trail of pheromone: The pheromone trail of the best ant is then updated to reinforce its path, encouraging other ants to follow it.

Check the end condition: The algorithm repeats the above steps until a stopping criterion is met.

End: The algorithm concludes when it satisfies the stopping condition and provides the best solution discovered.

2.5 Artificial bee colony optimization

The artificial bee colony (ABC) optimization imitates bee behaviour. A colony of bees is made up of onlookers, scouts and worker bees [ 47 ]. Artificial bees are flown in this system in a multidimensional search room and, depending on the experience they have gained and based on their next partner experience, the used bees pick their food sources and bees to change positions. Scout bees pick their food sources at random without any experience. Each food source avoids the probable solution to the problem under discussion [ 48 ]. The number of bees employed is as high as the food sources, each being a site, which is currently being used or as many solutions as individuals [ 49 ]. This procedure is continued until the ABC optimization meets a stop criterion.

ABC_Optimization (n, m, k)

population <- initialize (n, m, k)

global_best <- assign_random_food_source(population[m])

while! stop_criteria_met ()

for bee in population

fitness <- calculate_fitness (bee. food_source)

if fitness > global_best. fitness

global_best <- bee. food_source

for the bee in population

bee. update_food_source (global_best, bee. next_partner)

update_food_source (global_best, next_partner)

prob <- random_probability ()

if prob < experience

food_source <- global_best

else if prob < experience + next_partner. experience

food_source <- next_partner. food_source

food_source <- random_food_source ()

Initialization phase:

Initialize the x i j solution population in the j domain parameter. The exact description may be used for that purpose:

where x m a x j is the upper bound of the parameter j and x m i n j is the lower bound of the parameter j.

Worker bee phase:

Each worker bee uses a formula to identify and assess a food source v i j representative of a location such as a food source in her x i j memory. Each worker offers information about their food source to onlookers who select a food source website based on information collected from their bees while they wait at the hive according to Equation (2) :

If x k is a randomly picked solution, j is a parameter randomly selected and β i j is a random integer within the [–a, a] range. A greedy selection between v i and x i is applied after the production of a new solution v i ⁠ .

Onlooker bee phase:

There is a reference previously to the proportion of the amount of a food source to its location in the solution. Onlookers are positioned at food sources using a selection strategy based on fitness, such as the way of selecting the roulettes wheel. New solutions x i based on pi are picked to assess the new solutions v i and new solutions v i for spectators are created. The hired bees between v i and x i receive a greedy selection.

Scout bee phase:

Former workers who lost their resources start scouting randomly for food supplies. Every colony of bees has scout bees. The scouts have little instruction when looking for food. They mostly focus on finding food. Artificial bees can find the available answers rapidly. ABC decides that the artificial scout is the bee whose food supply has been lost or whose profitability has fallen below a specific level of profitability. The control parameter that decides the class is the withdrawal criterion or the ‘limit’. After a predefined number of attempts, a worker bee leaves an unimproved solution that is a source of food. The number of tests necessary to release the answer is determined by ‘limits’.

2.6 Particle swarm optimization

Particle swarm optimization (PSO) is a population-based evolutionary computational technique that is employed to address stochastic troubleshooting. It belongs to the category of swarm intelligence and is founded on social and psychological principles. PSO provides valuable insights into engineering applications and contributes to their development [ 50 ]. Social impact and social learning make cognitive consistency possible for the person. People may resolve issues by talking to people and by changing their ideas, attitudes and behaviour; they can usually be portrayed as people moving in a socio-cognitive space towards one another. But PSO has certain inconveniences such as global convergence; unlike some other optimization algorithms, PSO does not have a guarantee of global convergence, which means that it may not find the true optimal solution. To address this drawback, a novel PSO and a chaotic PSO are designed to tackle energy-system optimization issues efficiently. The analysis of the problem of unit commitment within the regulated system leads to the examination of UCP (uniform customs and practice for documentary credits) inside the deregulated market. The overall profit, execution time and convergence criteria are compared between various approaches.

One element is the current velocity of the particle v ( t ) ⁠ . Another is the optimum position Y ∗ ( t ) to approach the particle. The third factor is that the community or sub-community is optimally informed by Y ∗ ∗ ( t ) [ 51 ]. In each iteration step, the particle speed is modified to Y ∗ ( t ) and Y ∗ ∗ ( t ) ⁠ . Meanwhile, the random weight is independently allocated to the V i ⁠ , Y ∗ ( t ) and Y ∗ ∗ ( t ) ⁠ . The speed and position are updated following Equations (3) and (4):

In the given equation, v k +1( i , j ) represents the velocity of the particle in the i -th particle and j -th dimension at iteration k  + 1.

The weight factor ω determines the extent to which the previous velocity influences the new velocity.

v k ( i , j ) denotes the velocity of the particle in the i -th particle and the j -th dimension at iteration k .

C 1 and C 2 are the learning parameters that determine the influence of the personal best and global best solutions, respectively.

r and 1 and r and 2 are randomly generated numbers within the range of [0,1].

P bes t k ( i , j ) represents the personal best position of the i -th particle in the j -th dimension achieved thus far.

Y k ( I , j ) represents the current position of the i -th particle in the j -th dimension.

G b es t k signifies the global best position discovered by all particles up to the present iteration.

The flow chart for PSO is shown in Fig. 5 .

Flow chart of particle swarm optimization

2.7 Regression model

The research model [ 52 , 53 ] can be defined using Equation (5) :

where Y represents the dependent variable; this refers to the indication of respondent i ’s willingness to adopt smart home technology and their level of flexibility in terms of demand for technology j. β refers to the intercept. X 1 ij ,..., Xnij are dichotomous predictors included in the model. εij represents the random error term.

Building on Equation (5) , the level 2 model can be formulated as follows:

In Equations (5) and (6) , u 0 j ,..., u 1 j represent the random effects. W 1 j and W 2 j correspond to grand-mean centred and uncentred variables, respectively.

These equations are utilized in research to describe the relationships between the dependent variable, predictors and random effects. Equation (5) serves as the core model equation, capturing the influence of the predictors on the dependent variable while accounting for random error. Equations (5) and (6) extend the model by specifying the relationships and random effects associated with the intercept and predictor coefficients at the level 2 analysis. Collectively, these equations offer a comprehensive framework to analyse the variables that impact the acceptance of smart home technology and the adaptability of demand within the specific research context.

2.8 Regression and classification problems using AI

The RNN is a variation of the neural network frequently employed in the power systems domain to address regression and classification problems that involve sequential data. Unlike direct neural network models, the structure and operating principle of the RNN differ significantly [ 54 ]. In an RNN, the input data are fed to the model sequentially at each time step ( t ), as shown in the signal propagation diagram. At each step, the current state ( output ) is calculated by considering the current input data and the previously computed state. This iterative process continues for a fixed number of steps ( n ) until the desired output (predicted value) is achieved or until all input data ( input ) have been processed [ 55 ].

The propagation of signals in the RNN model is illustrated by the values assigned to each hidden state (hidden). These hidden states are calculated using the previous hidden state ( hidden  − 1) and current input data ( input ) [ 56 ]. hiddent = (〈 w hidden , hiddent – 1〉 + 〈 w input , input 〉) Here, σ () represents the activation function (such as the sigmoid function, hyperbolic tangent or rectified linear unit (ReLU)), while w hidden and w input are the weights for the hidden and input states, respectively.

The output value at each calculation step output is obtained by taking the dot product of the weights associated with the output state and the values of the hidden state, similar to a regression equation: output = 〈w output , hiddent 〉 [ 57 ].

During training, the initial stage involves calculating the output signal, after which the error function is calculated to determine the discrepancy. For regression problems, it is common to utilize the square root of the standard deviation between the output of the RNN and the values from the response space ( y t ):

Applying the chain rule, the gradient of the error functional is calculated. The weight coefficients ( w ij ) are adjusted in a manner that reduces the functional, following the direction of decreasing values, until it reaches the minimum value or the training iterations reach the predetermined limit. It is important to note that the weights associated with the hidden state of the RNN ( w hidden ) remain unchanged after propagating the error from each output ( output ). Conversely, the coefficients w output and w input change at each step of the gradient [ 58 ].

3.1 Operation of the power system

ELD is the process of assigning the generation output to the generation unit to supply the system load fully and economically. The whole engaged generating unit produces total electricity costs to minimize energy. The main concern of the ELD problem is to reduce the overall fuel cost. This is achieved by generating electricity in a way that optimizes the use of resources and reduces the overall cost of power production in the electricity system. Multiple generators provide enough total output to meet the consumer requirements in a typical power system. The production costs of each generating unit in the electricity system are different, as the producing units are not the same distance from the loading unit. Over the years, several AI approaches have been created to address this challenge. After simulation, authors concluded that using the genetic algorithm technique to solve the ELD problem can result in a lower overall cost of electricity production, but may also result in higher emissions. However, choosing a solution with a higher cost may result in minimum emissions. Transmission losses are usually neglected when they are small. However, for long-distance transmission in large and interconnected networks, transmission losses become significant and have an impact on the optimal distribution of power generation. It is possible to operate the same multitasking system with a better voltage profile and with evolutionary calculation technology, the cost and emission value of the best compromise.

The proposed results of the simulation of the ANN emphasize that the results are indiscernible from conventional methods, although the time used by neural networks is less than conventional methods. The number of generators increases the prediction error because there is a lot of input and output data to be learned. To monitor the performance, neural networks have been modelled. The authors have performed the simulations with many generation units having ramp rate limits and prohibited operating zones as constraints and the resultant performance is compared with ANN, GA and ACO techniques, but the ABC technique gives better outputs with fast convergence. The greedy selection procedure and the timely abandonment of the used food sources contained in ABC give it this potential. The basic operations of ABC optimization prevent solutions from stopping and make the algorithm more exploitative.

3.1.2 Generator maintenance scheduling

Generator maintenance scheduling (GMS) is a complicated combinatorial optimization issue for a power provider. Mathematical approaches include traditional ways to tackle the GMS issue. To evaluate the needed objective function, a mathematical model approach employs a trial-and-error procedure. Mathematical approaches even fail to provide viable answers as in some circumstances the operator needs to rely on certain assumptions and models that may not accurately reflect real-world conditions. In some cases, operators may need to be involved to provide additional input and expertise to make informed decisions about maintenance schedules. In addition, there may be unpredictable factors, such as equipment failures or changes in demand, that cannot be accurately accounted for by mathematical models alone. Maintenance is a preventive outage program for generating units within a certain time horizon in a power system. In the event of a range of various specification generating units in the energy system and several limitations to produce a sustainable and practical solution, maintenance planning becomes a difficult challenge. The maintenance planning of the generators is done for time horizons of different lengths. Short-term maintenance plans from 1 hour to 1 day are crucial to the daily operations, engagement and operational planning of power plants. Medium-term planning is necessary for resource management between 1 day and 1 year. Long-term planning from 1 to 2 years is crucial for future planning. An examination is being conducted to resolve some AI methods, including simulated developments, neural networks and GAs. The application of the genetic algorithm through case research shows that suitable GA parameters are safeguarded, as well as issue coding and development functions. The use of integer encoding decreases the velocity of the genetic search method investigation. By using integer encoding, the algorithm needs to perform additional operations to convert solutions into integers, which can slow down the search process. Planning the generation of power remains a barrier to competent solution technology and a difficult optimization problem. The challenge in power generation planning lies in finding the optimal balance between cost and efficiency, while also considering factors such as environmental impact, reliability and security of supply. The answer to the difficulties in generation planning consists of finding the UC (unit commitment) at every point in the programming period for each generator in one power system. An electrical system must be defined in each planning interval for each power generator for the decisions and levels of output. The solution process must be addressed concurrently for binary decisions and continuous variables. Generation difficulties with scheduling are typically quite narrow and combined. Match swarm optimization approaches have been used to achieve viable schedules within a specified time. The study found that an optimization-based approach using PSO provided better results than a GA or an evolutionary strategy. Data from the actual power system were used to evaluate the performance of the different optimization techniques.

UC is properly scheduled for the ON/OFF status and the genuine generator power outputs of the system. To satisfy a high number of system limitations and decrease the overall fuel cost at every time interval, a spinning reserve is necessary (spinning reserve refers to the additional generation capacity that is available and running but not actively supplying power to the grid). UC meets the expected load requirements in advance. To implement UCs, medium-term load forecasting using ANN was used. The neural network structure was trained through learning and parameter learning. Total operational expenses under 24 hours were used for the assessment criterion. The study demonstrates the effectiveness of the proposed approach by comparing the performance of the ANN-based load forecasting model with traditional methods such as linear regression and time-series forecasting. The results show that the ANN-based load forecasting model significantly improves the accuracy of load forecasting and reduces the scheduling cost by reducing the number of units needed for scheduling. The study also highlights the importance of considering the uncertainty and variability of load demand in UC scheduling and suggests that ANN-based load forecasting models can be a useful tool for achieving more efficient and reliable scheduling in power systems. Locational marginal prices have been evaluated through a trained ANN. The findings show that the current technique gives a different UC mechanism. To develop unit commitment, the PSO technique is used. On implementation, with the increasing size, the execution time is also increasing. To accelerate the PSO, a convergence repair method is also implemented.

3.1.4 Optimal power flow

Optimal power flow (OPF) is a highly important technique to identify the optimum control parameter settings that enhance or decrease the intended target function, but also under a variety of limitations. An essential instrument to design and operate a power system is the issue of optimum power flow to identify the best parameter settings that can maximize or minimize the intended goal function within specific limitations. Voltage and reactive controls, also called OPD, are an OPF sub-problem that seeks to reduce overall transmission loss by resuming the reactive power flow. Optimal reactive power dispatch is a non-linear solution for the issue of blending integers since some control variables such as tap ratios for transformers, shunt capacitor outputs and reactors are distinct.

The alternate strategy for mitigating the problem of GA-ANN is set out in this article. A collection of ANN networks is trained offline in specified system quantities to work on a general OPF issue. To choose the appropriate ANN inputs, the k-mean clustering technique is utilized. When learning the functions correctly, ANNs can easily estimate the associated results with great precision.

The ANFIS develops the input/output data set fuzzy-inference system (FIS) that matches the membership (adjustment) parameters with a backpropagation or minimum square process type. This update allows you to learn from the fuzzy systems data IEEE 39 bus system implementations and simulated software from the power world are utilized for the formation of the ANFIS. The results indicate that the ANFIS offers solutions as accurate as conventional ones. It takes less time, though, and it works quickly. Some additional papers on the application of AI in the operation of power systems are presented in Table 3 .

Applications of artificial intelligence in the operation of a power system

3.2 Control of the power system

The main objective of a voltage controller power system is to maintain the voltage profile within a defined limit, thus minimizing transmission losses and avoiding cases of voltage instability [ 81 ]. The VC system consists of three levels of hierarchical control: AVR (automatic voltage regulator), tertiary voltage control (TerVC) and secondary voltage control (SecVC). AVR is aimed at controlling the voltage of buses that are equipped with reactive power sources (e.g. synchronous, sync, static var compensators and STATCOM (static synchronous compensator)). Actions are carried out locally at this control level. SecVC is used to monitor the voltage on a specific bus that controls a cargo bus.

In situations in which there is hardware present in the vicinity that modifies the reference point for the AVR, the control level typically operates at a slower pace compared with the AVR control level. SecVC is responsible for identifying VC regions and their correlation with individual load buses. To accommodate varying power system conditions, SecVC must demonstrate flexibility in adjusting the control regions to accommodate all grid conditions. On the other hand, TerVC determines the optimal reference value for voltage grids at each load bus. The objective is to minimize power loss, optimize reactive power and maintain a minimum load release or reservation. TerVC is usually updated every 30 minutes to 1 hour.

The backward error propagation algorithm trains the multilayered feedforward perception. The minimum singular value method analyses the static voltage collapse. The procedure uses a minimum voltage stability evaluation time once the network training is complete. For monitoring voltage collapse, complementary methodologies of neural networks and expert systems would be combined for use in the application [ 82 ].

GA is an iterative optimization technique with several solutions from the candidates (known as a population). In the case in which there is no knowledge of the problem field, then the GA starts to look for solutions from a random population. The appropriate coding (or display) must first be defined to solve the problem. A fitness function should also be defined so that every coding solution is given a figure of merit. If parents are not satisfied with the termination condition, for reproduction, they must be picked [ 83 ]. They are then joined to generate offspring through reproduction and, to refresh the population of candidate solutions, crossover and mutation operators are utilized. Typically, in a basic genetic algorithm, three operators are involved: selection, crossover and mutation. These operations are performed to generate new offspring, individuals and subsequent generations. The same process is repeated with the new generation until the desired criteria are met. The approach of this method is used to teach swarming at the beginning of PSO. In this case, 10 control variables are used for the ANN input. The neuron and its prejudice are 11. A hidden layer consists of this group of neurons and biases. Ten outputs/goals are available. These objectives are achieved by using the optimal value of PSO. The last outputs are the initializations in the time-varying non-linear particle swarm optimization (TVNL-PSO). The steps are as follows:

PSO is used to take the ANN input; the weight value is applied at random;

the ANN input and partition in a cached layer are weighed and then activated by the sigmoid binary function;

weighting of the output in the hidden layer and activation of the linear function;

to optimize the reactive power and VC by TVNL-PSO, the ANN output is transmitted as a starting initialization value.

3.2.2 Power system stability control

The stability of a power system is a feature that allows it to remain under a balance in normal operating conditions and retrieve an acceptable balance after a change. Margins of stability can be seen to decline throughout the world [ 84 ]. We highlight three of the many reasons for this:

The inhibition by economic and environmental constraints of further transmission or construction. Therefore, power systems must be operated with lower safety margins.

Restructuring of the electricity industry. The restructuring process reduces the margins of stability, as power systems do not co-operate effectively [ 85 ].

Increased complexity of power systems multiplies the compulsive properties. These include large, non-linear oscillations; frequency differences between weakly binding energy-system areas; interactions with saturated devices.

Fuzzy logic endeavours to address problems by emulating human reasoning, allowing optimal decision-making based on available information. It can also be employed to regulate the stability of un-modelled systems. To achieve improved performance, a fuzzy logic (FL) controller is combined with a PID (proportional–integral–derivative) controller. In this particular scenario, the fuzzy logic control adjusts the gains of the PID controller based on the power system.

A fuzzy logic controller primarily consists of four major parts: fuzzification, fuzzy rule base, fuzzy inference and defuzzification. FACTS have proven to be extremely promising for increasing performance under stable conditions. The most promising FACTS device is a unified power flow controller (UPFC). Three control factors can be adjusted: bus voltage, reaction line and phase angle between two buses. The power flow should be redistributed across lines while a stable state is preserved. It can also be utilized to increase damping when low frequencies are damped temporarily.

Power system stability control.

Load frequency control as defined by the controllable generator power output control in a prescribed area resolves system frequency changes, two-line loadings or interactions to maintain an interchange with other regions within the fixed limit or scheduled system frequency [ 86 ]. The traditional proportional–integral (PI) controller is the most widely used among different types of load frequency controllers. The PI controller can be easily implemented and provides a faster response, but its performance decreases when unwanted disturbances, such as load change dynamics, increase the difficulty within the system. In this paper, less computing is required for the non-linear autoregressive-moving average-L2 (NARMA-L2) control architecture. Plant output, reference and control signals are included. The controller is therefore taught to monitor the output of the reference model. The model network that updates controller settings predicts the effect of the change in plant performance. Some additional papers on the application of AI in the control of power systems are presented in Table 4 .

Applications of AI in control of a power system

3.3 Planning of power system

DSP plays a crucial role in enhancing reliability and minimizing costs for both utilities and consumers. Electric power distribution networks are a fundamental component of the electrical power system. In general, transport networks are denser and more complex than those that provide transformer stations [ 110 ]. Automating previously manual jobs increases with distribution networks becoming more complex. New tools are known as advanced automation functions that support the operation of such networks. These functions enable the network operator to effectively address issues that arise. Furthermore, the reconfiguration of distribution networks is essential to identifying optimal solutions that align with the operator’s requirements and constraints, ensuring a secure and economically optimized electricity supply.

The optimal design of a distribution network is not a fixed solution, but rather a process that involves considering various technically feasible options and using improvement tools to make the best decision based on factors such as demand, reliability of power transmission and network structure. All potential paths are initially identified with uploaded system data and then the energy-loss cost calculation applies for each identified path forward/backward sweeping load flow technique. For the distribution of power, the minimum energy-loss path is chosen. The optimal selection of the branch conductor of the radial system is done using optimization of PSO. In this case, parameters such as power loss, voltage profile and capital investment depreciation improve optimization. These parameters are used as optimization criteria to determine the optimal branch conductor that minimizes power loss, improves the voltage profile and reduces capital investment depreciation. The PSO algorithm iteratively updates the position of each particle in the search space based on its own experience and the experiences of its neighbours. The algorithm continues until a global optimum is found or a stopping criterion is met. The final solution produced by the PSO algorithm represents the optimal branch conductor that meets the optimization criteria [ 111 ]. The optimization of the PSO results in the optimal conductor and the best substation, the positioning of the optimal conductor is selected and then the optimal substation power distribution is achieved.

A multi-target algorithm was proposed that uses a fluid optimization technique to handle contradictory targets [ 112 ]. The plan formulation and the algorithm include a multi-target function that uses battery energy storage systems (BESSs) and traditional resources to select the best planning option. The microgrid BESS has been receptive to power management and improvement in power quality. The proposed algorithm is based on the fuse-based decision-making processes of the Mamdani-type FIS and Bellman–Zadeh approach.

In this paper [ 113 ], two algorithms, namely the mixed-integer linear program (MILP) and GA, are compared for the design of a radial distribution system feeder. The main objective is to minimize total investment and operational outages while maximizing system reliability. The study aims to evaluate and compare these two optimization techniques in terms of their optimality, complexity and time requirements. A unique aspect of the optimization model is the consideration of operational costs associated with failures, which are directly linked to the design of the system. The fault rate and defect cost at each loading point are updated based on the proposed configurations. It is crucial to determine which method produces superior results in terms of optimality, complexity and time efficiency.

The GA technique is used to build the algorithm for optimizing distribution networks. The fundamental concept is the growth of the genetic operator population (selection, crossover and mutation). These are used to generate a fresh population from the previous generation throughout each generation procedure. In GA, a single chromosome shows each person. This chromosome corresponds, according to the graph theory, to a radial distribution network configuration or a particular graph twist. The chromosome group is the population. Randomly, an initial group is created as a first step in the implementation of GA. Then the encoding is applied to each chromosome. In this study, only closed branches represent the network topology. A true coding strategy was used to match each gene to the edge of the coagulation tree [ 114 ].

3.3.2 Reactive power optimization

As the demand of electricity increases and new lines are built, the environment and the unforeseen power fluid in the lens are reduced, it is generated by the current scenario. Effect reactive compensation control improves voltage, reduces energy loss and improves system performance under stable and dynamic conditions in weak nodes [ 115 ]. Because the complexity of power systems is constantly increasing and the network components are constantly being loaded, abnormal operating conditions such as voltage can occur more often. Therefore, it is obvious that the power system needs adequate reactive power and VC.

In pursuit of intelligent theory development, a combination of fuzzy logic and ANNs is used to determine the control strategy for transformer taps and capacitors. However, due to the increasing complexity of control variables, rapid optimization becomes challenging [ 116 ]. To address this issue, a genetic algorithm is utilized, which tackles problems associated with incorporating regulatory time as penalty terms in the objective function and determining appropriate penalty factors that affect algorithm performance. When regulatory time is a constraint, the optimization objective focuses on minimizing the total energy loss during the dispatch period. GAs, inspired by natural selection mechanics and genetics, such as inheritance, mutation and recombination, are utilized (also referred to as crossover) to optimize the solution.

The PSO method can be used for handling FACTS devices in power systems. Various reactive power problem objectives and different solutions are addressed in the interconnected power system. Solutions and comparative analyses using the FACTS device, differential evolution (DE) and PSO algorithms are presented under various loading conditions [ 117 ]. The algorithm proposed [ 118 ] employs DE to minimize generator fuel costs on FACTS devices. Additionally, the authors discuss the hybridization of DE and PSO (DEPSO) to overcome the maximum load limit. The control of reactive flow is addressed using fuzzy sets and a fuzzy feature optimization technique is introduced to optimize reactive power. The utilization of fuzzy linear programming offers an effective approach to calculating reactive power, aiming to minimize active power loss and maximize the voltage stability margin. The paper [ 119 ] explores the combination of fuzzy and GA approaches for FACTS shunt controller placement and sizing. Lastly, the focus of [ 120 ] is on the integration of fuzzy systems with GM algorithms and the PSO algorithm to tackle the OPF problem and optimize control variables. In this paper, the authors [ 121 ] focus on fluid-based reactive and voltage controls to reduce actual loss of power.

3.3.3 Capacitor placement

There are some advantages if capacitors are placed optimally, including various factors, such as maximizing energy and reducing peak power loss through the introduction of a condenser in an electrical distribution system. In the paper, a novel adaptive modified firefly algorithm is presented to address the optimal capacitor placement problem in power systems. This optimization problem involves identifying the best positions and sizes of capacitors in a power system, in order to enhance voltage stability, minimize energy losses and improve the power factor. The proposed algorithm combines the firefly algorithm with adaptive parameter settings and introduces a unique crossover operator to enhance both convergence speed and solution quality. The authors conduct evaluations on a test system and compare the performance of the algorithm with other optimization methods. The results demonstrate the effectiveness of the proposed approach in finding optimal solutions and highlight its superiority in terms of solution quality and computational efficiency compared with other algorithms [ 122 ]. Losses occur due to reactive currents in the distribution system and are therefore minimized in the right places. Shunt capacitors are used, depending on their use. A capacitor is used to improve the voltage profile, reduce losses and increase the power factor [ 123 ].

Elbaz et al. [ 124 ] have been using ANN techniques to control both capacitor banks and voltage regulators. The ANN has many input connections and all inputs are combined to determine the output capacity. The purpose of the capacitor search algorithm was to reduce total active losses in the distribution system by utilizing the capacitor banking search to address the capacitor placement problem. The operation of the ant colony was proposed to address problems related to the installation of the capacitor. The fuzzy method uses variables such as angle, current and voltage, etc. A degree for a set and fuzzy variable is determined by MFs. This degree changes from zero to one that takes zero or one as opposed to the classical methods [ 125 ]. A fuzzy logic-based algorithm is developed to minimize line loss for the placement of condensers in a radial system. The fuzzy expert system identifies the capacitor candidate nodes by compromising the possible reduction in loss between the condenser system and the voltage level. This paper [ 126 ] presents a fluffy approach to identifying the appropriate places for capacitor placement. In the design of a seamless logic to determine the optimal placement of capacitors, two main objectives are taken into account. These objectives include (i) minimizing actual power loss and (ii) maintaining voltage within acceptable limits. Fuzzy member functions are used to model voltage and power loss indices for nodes in the transmission system [ 127 ]. The suitability of capacitor placement for each node in the distribution system is then determined. This is achieved using a set of rules within the FIS. Nodes can be fitted with capacitors of greatest appropriateness.

In [ 128 ], a method based on GAs is used to identify the optimal locations and sizes of capacitors in the distribution network. The capacitor sizes are considered discrete and known variables placed on the buses of the network. Hence, the maximum losses of the distribution system are reduced. GA technology is selected as the capacitor problem is more accurately addressed in the power grid. When the search area crosses for an optimal solution, the advantages of GA are multifunctional—that is, when a locally optimal solution is found to an engineering goal, GA adapts its search to find an optimal global solution, subject to predefined search restrictions [ 129 , 130 ]. The article shows the results of the study of the best size and location for a GA-connected system using bays in Saudi Arabia in West–East regions [ 131 , 132 ]. Two formulas are proposed for capacitor positioning: (i) cost balance condenser/loss principle and (ii) total system performance cost estimates, standard analysis and verification of annual benefits, power loss and operational tension results [ 133 , 134 ]. AI is applicable for various aspects of the power system [ 135 ]. Some additional papers on power system application of AI in planning are presented in Table 5 .

Applications of AI in the planning of a power system

Application of AI in various problems of the operation, control and planning of power systems has shown good performance over conventional methods. ABC algorithms function better than other AI techniques in the ELD problem, as shown in a comparison. The ABC algorithm has the highest-quality solutions, consistent convergence and exceptional computing efficiency. Compared with traditional mathematical models, both GA and PSO techniques are superior; the PSO technique is preferable since GA replaces humans. Unlike GA, PSO models alter humans through time and, in the following generation, all people survive. The fate of each person is continually adjusted based on the global ideal point thus far. The swarming effect in PSO allows the population of particles to move collectively in the search space, facilitating a more efficient exploration of the solution space and a faster convergence to optimum search areas compared with GA. PSO is particularly useful for optimization problems with many variables and when the solution space is complex and not well understood. Additionally, PSO requires minimal computational resources compared with GA as it does not require the evaluation and selection of multiple generations of offspring. In UC, with the increasing time, the execution time is also increasing, so the combination of many AI techniques can be progressive, as it can potentially increase the efficiency and accuracy of the optimization problem. The scalability of any heuristic optimization method is a major issue. In the load flow method, the ANFIS and ABC algorithms provide efficient and accurate solutions, and the implementation of ANNs is fast and can handle missing data effectively. In VC, rather than replacing conventional methods, the focus should be on enhancing their capabilities through the integration of AI techniques, modern control theory, fuzzy technology and ANNs, along with adaptive control and expert systems. This approach, which combines current research trends with practical experience, has great potential for practical applications. In load frequency control, the NARMA-L2 ANN network architecture provides the best output after some series of trials and improvements. Under fault conditions, it has been observed. The fuzzy-PID power output works so that the power systems are fast and stable. Due to the specialization, the fuzzy logic in condenser placement is better than other approaches. Fuzzy logic includes the relatively basic technique of calculating the necessary nodes in the distribution system to replace condensers with approximate reasoning. The results of the study indicate that the GA method can provide a globally optimal solution for reactive power optimization, particularly when an ample generation and population size is used. Furthermore, it was found that the use of the UPFC resulted in minimized losses compared with the static var compensator and thyristor-controlled series capacitor. In terms of reactive power planning, a fuzzy membership-based approach is employed in an interconnected power system to identify weak nodes and determine optimal parameter settings for FACTS devices. The feasibility of this strategy is validated through its solutions and compared with other global optimization approaches. The proposed technique is applied to a standard system under high load conditions, resulting in a stable system with reduced losses and cost savings. This approach has the potential to become a novel technology for effectively coordinating FACTS devices with other existing generators. Following the pandemic, it is expected that governments around the world will prioritize energy efficiency in buildings and smart homes. To support this, the development of open-source protocols and unified connectivity solutions is crucial. Smart home systems are focused on maximizing the energy efficiency of major household appliances, thereby contributing to overall energy balance. By constructing sustainable homes integrated with smart technologies and a combination of energy sources, significant cost savings and a reduction in carbon footprints can be achieved. While smart homes are becoming more common, there are still barriers to widespread adoption that researchers need to address. It emphasizes the potential of smart home companies and highlights technical challenges such as device compatibility. The review also discusses the importance of studying consumer attitudes and demands, and mentions the limitations of the survey methodology. Significant scientific results include an algorithm for a modern predictive analytics system, an approach to assess the condition index of the equipment and a method to determine the probability of defects using ML. The study validated the model using data from a hydroelectric power plant, demonstrating its accuracy. Future research can focus on refining the index calculation for equipment with multiple functional units and constructing predictive models for specific equipment classes. ML and data-driven techniques hold great promise in the field of power systems, especially in the context of smart grids. These methods can effectively analyse and process large volumes of data, resulting in improved accuracy and increased operational efficiency. However, some challenges need to be addressed, such as ensuring the quality of the data and interpreting the results in a meaningful way. Accurate forecasting plays a vital role in optimizing grid operations, integrating renewable energy sources and achieving cost-effective power generation. ML plays a vital role in transforming traditional grids into smart grids, improving reliability and safety. It also aids in demand-side management and enhances cybersecurity.

Evaluating energy costs and making improvements can lead to significant energy savings. Smart technologies have the potential to reduce electricity demand and environmental impact. Social acceptance of smart home systems needs to be promoted. Further research can expand sample sizes, include more diverse countries and explore smart meter readings. Future research should emphasize the importance of addressing technical, security and privacy concerns, and call for collaboration between stakeholders to enhance the smart home market. Although the developed approach offers several benefits, there is still an unresolved issue regarding the calculation of the technical condition index for equipment consisting of multiple functional units. Existing methods rely on assigning weights to each unit based on an expert evaluation to determine its importance. Additional research can focus on improving the methods for calculating the technical condition index for different types of power equipment and establishing predictive models to anticipate equipment defects in the event of functional unit failures. Future investigations should prioritize the development of more precise and dependable predictive models for power systems, taking into account the challenges related to data availability and interpretability.

Anshumaan Pathak and Utkarsh Pandey did the critical review. Surajit Mondal and Adesh Kumar supervised and reviewed the manuscript.

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

The paper is a study and review article in which no specific data are referred to. No simulation software is used.

Sozontov A , Ivanova M , Gibadullin A. Implementation of artificial intelligence in the electric power industry . E3S Web of Conferences , 2019 , 114 : 01009 .

Google Scholar

Zhao X , Zhang X. Artificial intelligence applications in power system . In: Luo X (ed). Advances in Intelligent Systems Research—2016 2nd International Conference on Artificial Intelligence and Industrial Engineering (AIIE 2016). Dordrecht, Netherlands : Atlantis Press , 2016 , 158 – 161 .

Nath RP , Balaji VN. Artificial intelligence in power systems . IOSR Journal of Computer Engineering (IOSR-JCE) , 2014 . https://jcboseust.ac.in/electrical/images/notes/aitech_ug_ai_reactive_power_control.pdf ( 29 July 2023 , date last accessed).

Das S , Mukherjee M , Mondal S. Detailed energy audit of thermal power plant equipment . World Scientific News , 2015 , 22 : 70 – 90 .

Mondal S , Mondal AK , Sharma A , et al.  . An overview of cleaning and prevention processes for enhancing efficiency of solar photovoltaic panels . Current Sci , 2018 , 115 : 1065 – 1077 .

Wasesa M , Tiara AR , Afrianto MA , et al.  . SARIMA and artificial neural network models for forecasting electricity consumption of a microgrid based educational building . In: 2020 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM), Singapore, 14–17 December 2020 , 210 – 214 .

Jang JS. ANFIS: adaptive-network-based fuzzy inference system . IEEE Trans Syst Man Cybern , 1993 , 23 : 665 – 685 .

Borges AF , Laurindo FJ , Spínola MM , et al.  . The strategic use of artificial intelligence in the digital era: systematic literature review and future research directions . Int J Inf Manage , 2021 , 57 : 102225 .

Guan Y , Chaffart D , Liu G , et al.  . Machine learning in solid heterogeneous catalysis: recent developments, challenges and perspectives . Chem Eng Sci , 2022 , 248 : 117224 .

Bordin C , Skjelbred HI , Kong J , et al.  . Machine learning for hydropower scheduling: state of the art and future research directions . Procedia Comput Sci , 2020 , 176 : 1659 – 1668 .

Wang H , Ricardez-Sandoval LA. Dynamic optimization of a pilot-scale entrained-flow gasifier using artificial recurrent neural networks . Fuel , 2020 , 272 : 117731 .

Isnen M , Kurniawan S , Garcia-Palacios E. A-SEM: an adaptive smart energy management testbed for shiftable loads optimisation in the smart home . Measurement , 2020 , 152 : 107285 .

Bibri SE , Krogstie J. Environmentally data-driven smart sustainable cities: applied innovative solutions for energy efficiency, pollution reduction, and urban metabolism . Energy Informatics , 2020 , 3 : 1 – 59 .

Strielkowski W. Social Impacts of Smart Grids: The Future of Smart Grids and Energy Market Design . Amsterdam, Netherlands : Elsevier , 2019 .

Google Preview

Oliveira L , Mitchell V , May A. Smart home technology—comparing householder expectations at the point of installation with experiences 1 year later . Pers Ubiquitous Comput , 2020 , 24 : 613 – 626 .

Tao M , Zuo J , Liu Z , et al.  . Multi-layer cloud architectural model and ontology-based security service framework for IoT-based smart homes . Future Gener Comput Syst , 2018 , 78 : 1040 – 1051 .

Shcherbatov IA , Turikov GN. Determination of power engineering equipment’s defects in predictive analytic system using machine learning algorithms . J Phys Conf Ser , 2020 , 1683 : 042056 .

Moleda M , Momot A , Mrozek D. Predictive maintenance of boiler feed water pumps using SCADA data . Sensors , 2020 , 20 : 571 .

Ren C , Xu Y. A fully data-driven method based on generative adversarial networks for power system dynamic security assessment with missing data . IEEE Trans Power Syst , 2019 , 34 : 5044 – 5052 .

Yu J , Guo Y , Sun H. Testbeds for integrated transmission and distribution networks: generation methodology and benchmarks . CSEE J Power Energy Syst , 2020 , 6 : 518 – 527 .

Gonen T. Electric Power Distribution Engineering . Boca Raton, FL, USA : CRC Press , 2015 .

Wood AJ , Wollenberg BF , Sheblé GB. Power Generation, Operation, and Control . Hoboken, NJ, USA : John Wiley & Sons , 2013 .

Kumar N , Mishra VM , Kumar A. Smart grid and nuclear power plant security by integrating cryptographic hardware chip . Nuclear Engineering and Technology , 2021 , 53 : 3327 – 3334 .

Kumar N , Mishra VM , Kumar A. Smart grid security by embedding s-box advanced encryption standard . Intelligent Automation &. Soft Comput , 2022 , 34 : 623 – 638 .

Xu G , Wang Z. Power system load flow distribution research based on adaptive neuro-fuzzy inference systems . In: 2012 Spring Congress on Engineering and Technology, Xi’an, China, 27–30 May 2012 , 1 – 4 .

Karaboga D , Akay B. A comparative study of artificial bee colony algorithm . Appl Math Comput , 2009 , 214 : 108 – 132 .

Nualhong D , Chusanapiputt S , Phomvuttisarn S , et al.  . Diversity control approach to ant colony optimization for unit commitment problem . In: 2004 IEEE Region 10 Conference TENCON 2004, Chiang Mai, Thailand, 24 November 2004 , 488 – 491 .

Ho SL , Yang S , Ni G , et al.  . A particle swarm optimization-based method for multiobjective design optimizations . IEEE Trans Magn , 2005 , 41 : 1756 – 1759 .

Ratnaweera A , Halgamuge SK , Watson HC. Self-organizing hierarchical particle swarm optimizer with time-varying acceleration coefficients . IEEE Trans Evol Comput , 2004 , 8 : 240 – 255 .

Kumar N , Nangia U , Sahay KB. Economic load dispatch using improved particle swarm optimization algorithms . In: 2014 6th IEEE Power India International Conference (PIICON), Delhi, India, 5–7 December 2014 , 1 – 6 .

Panta S , Premrudeepreechacharn S. Economic dispatch for power generation using artificial neural network ICPE’07 conference in Daegu, Korea . In: 2007 7th International Conference on Power Electronics, Daegu, South Korea, 22–26 October 2007 , 558 – 562 .

Rahmat NA , Musirin I. Differential evolution ant colony optimization (DEACO) technique in solving economic load dispatch problem . In: 2012 IEEE International Power Engineering and Optimization Conference, Melaka, Malaysia, 6–7 June 2012 , 263 – 268 .

Saxena A , Pandey SN , Srivastava L. Congestion management in open access—a review . International Journal of Science, Engineering and Technology Research , 2013 , 2 : 922 – 930 .

Emami H , Sadri JA. Congestion management of transmission lines in the market environment . International Research Journal of Applied and Basic Sciences , 2012 , 3 : 2572 – 2580 .

Ghosh K , Pandey U , Pathak A , et al.  . Simulation of density based traffic control system using Proteus 7.1 professional . In: Mathur G , Bundele M , Tripathi A , Paprzycki M (eds). Proceedings of 3rd International Conference on Artificial Intelligence: Advances and Applications: ICAIAA 2022. Singapore: Springer Nature Singapore , 2023 , 493 – 504 .

Deb S , Goswami AK. Congestion management by generator rescheduling using artificial bee colony optimization technique . In: 2012 Annual IEEE India Conference (INDICON), Kochi, India, 7–9 December 2012 , 909 – 914 .

Choudekar P , Sinha SK , Siddiqui A. Transmission line efficiency improvement and congestion management under critical contingency condition by optimal placement of TCSC . In: 2016 7th India International Conference on Power Electronics (IICPE), Patiala, India, 17–19 November 2016 , 1 – 6 .

Wang X , McDonald JR. Modern Power System Planning . London, UK : McGraw-Hill , 1994 , 247 – 307 .

Dahal K , McDonald J. A review of generator maintenance scheduling using artificial intelligence techniques . In: Proceedings of the 32nd Universities Power Engineering Conference (UPEC ’97), Manchester, UK, 10–12 September 1997 , 787 – 790 .

Koay CA , Srinivasan D. Particle swarm optimization-based approach for generator maintenance scheduling . In: Proceedings of the 2003 IEEE Swarm Intelligence Symposium. SIS’03 (Cat. No. 03EX706), Indianapolis, IN, USA, 26 April 2003 , 167 – 173 .

Johnson RC , Happ HH , Wright WJ. Large scale hydro-thermal unit commitment: method and results . IEEE Trans Power Appar Syst , 1971 , PAS-90 : 1373 – 1384 .

Thum YM. Hierarchical linear models for multivariate outcomes . Journal of Educational and Behavioral Statistics , 1997 , 22 : 77 – 108 .

Reyes DM , de Souza RM , de Oliveira AL. A three-stage approach for modeling multiple time series applied to symbolic quartile data . Expert Syst Appl , 2022 , 187 : 115884 .

Almaghrebi A , Aljuheshi F , Rafaie M , et al.  . Data-driven charging demand prediction at public charging stations using supervised machine learning regression methods . Energies , 2020 , 13 : 4231 .

Huang P , Copertaro B , Zhang X , et al.  . A review of data centers as prosumers in district energy systems: renewable energy integration and waste heat reuse for district heating . Appl Energy , 2020 , 258 : 114109 .

Sun L , You F. Machine learning and data-driven techniques for the control of smart power generation systems: an uncertainty handling perspective . Engineering , 2021 , 7 : 1239 – 1247 .

Lisin E , Shuvalova D , Volkova I , et al.  . Sustainable development of regional power systems and the consumption of electric energy . Sustainability , 2018 , 10 : 1111 .

Prakash S , Jain J , Hasnat S , et al.  . Economic load dispatch with valve point loading effect using optimization techniques . In: Bansal P , Tushir M , Balas VE , Srivastava R (eds). Proceedings of International Conference on Artificial Intelligence and Applications: ICAIAA 2020, Singapore: Springer , 2021 , 407 – 416 .

Chowdhury AK , Mondal S , Mukherjee M , Biswas PK. Mega watt security assessment of power systems . International Letters of Chemistry, Physics and Astronomy, 2015 , 58 : 9 – 15 .

Dommel HW , Tinney WF. Optimal power flow solutions . IEEE Transactions on Power Apparatus and Systems, 1968 , PAS-87 : 1866 – 1876 .

Wakiru JM , Pintelon L , Muchiri P , et al.  . A comparative analysis of maintenance strategies and data application in asset performance management for both developed and developing countries . International Journal of Quality & Reliability Management , 2022 , 39 : 961 – 983 .

Muñoz-Delgado G , Contreras J , Arroyo JM , et al.  . Integrated transmission and distribution system expansion planning under uncertainty . IEEE Trans Smart Grid , 2021 , 12 : 4113 – 4125 .

Chowdhury AK , Mondal S , Alam SM , et al.  . Voltage security assessment of power system . World Scientific News , 2015 , 21 : 36 – 50 .

Holen AT , Botnen A , Stoa P , et al.  . Coupling between knowledge-based and algorithmic methods . Proc IEEE , 1992 , 80 : 745 – 757 .

Short MJ , Hui KC , Ekwue AO , et al.  . Applications of artificial neural networks for NGC voltage collapse monitoring . International Conference On Large High Voltage Electric Systems, 1994 , 2 : 38 – 205 .

Haida T , Akimoto T. Voltage optimization using genetic algorithm . In: Proc. 3rd Symposium on Expert System Applications to Power Systems, Tokyo, Japan, 1991 , 375 – 380 .

Denny FI , Dismukes DE. Power System Operations and Electricity Markets . Boca Raton, FL, USA : CRC Press , 2017 .

Masiala M , Ghribi M , Kaddouri A. An adaptive fuzzy controller gain scheduling for power system load-frequency control . In: 2004 IEEE International Conference on Industrial Technology, IEEE ICIT’04, Vol. 3 , Hammamet, Tunisia, 8–10 December 2004 , 1515 – 1520 .

Alquthami T , Butt SE , Tahir MF , et al.  . Short-term optimal scheduling of hydro-thermal power plants using artificial bee colony algorithm . Energy Rep , 2020 , 6 : 984 – 992 .

Sahay KB , Sonkar A , Kumar A. Economic load dispatch using genetic algorithm optimization technique . In: 2018 International Conference and Utility Exhibition on Green Energy for Sustainable Development (ICUE), Phuket, Thailand, 24–26 October 2018 , 1 – 5 .

Mishra SK , Mishra SK. A comparative study of solution of economic load dispatch problem in power systems in the environmental perspective . Procedia Comput Sci , 2015 , 48 : 96 – 100 .

Dixit GP , Dubey HM , Pandit M , et al.  . Artificial bee colony optimization for combined economic load and emission dispatch . In: International Conference on Sustainable Energy and Intelligent Systems (SEISCON 2011), Chennai, India, 20–22 July 2011 , 340 – 345 .

Daniel L , Chaturvedi KT , Kolhe ML. Dynamic economic load dispatch using Levenberg Marquardt algorithm . Energy Procedia , 2018 , 144 : 95 – 103 .

Ruiz-Abellón MC , Fernández-Jiménez LA , Guillamón A , et al.  . Integration of demand response and short-term forecasting for the management of prosumers’ demand and generation . Energies , 2020 , 13 : 11 .

Ali SS , Choi BJ. State-of-the-art artificial intelligence techniques for distributed smart grids: a review . Electronics , 2020 , 9 : 1030 .

Fu J , Nunez A , De Schutter B. A short-term preventive maintenance scheduling method for distribution networks with distributed generators and batteries . IEEE Trans Power Syst , 2020 , 36 : 2516 – 2531 .

Esmaili M , Shayanfar HA , Moslemi R. Locating series FACTS devices for multi-objective congestion management improving voltage and transient stability . Eur J Oper Res , 2014 , 236 : 763 – 773 .

Suresh K , Kumarappan N. Hybrid improved binary particle swarm optimization approach for generation maintenance scheduling problem . Swarm Evol Comput , 2013 , 9 : 69 – 89 .

Lakshminarayanan S , Kaur D. Optimal maintenance scheduling of generator units using discrete integer cuckoo search optimization algorithm . Swarm Evol Comput , 2018 , 42 : 89 – 98 .

Scalabrini Sampaio G , Vallim Filho ARDA , Santos da Silva L , et al.  . Prediction of motor failure time using an artificial neural network . Sensors , 2019 , 19 : 4342 .

Fikri M , Cheddadi B , Sabri O. Predicting Moroccan real network’s power flow employing the artificial neural networks . In: 2019 Third International Conference on Intelligent Computing in Data Sciences (ICDS), Marrakech, Morocco, 28–30 October 2019 , 1 – 6 .

Rahul J , Sharma Y , Birla D. A new attempt to optimize optimal power flow-based transmission losses using genetic algorithm . In: 2012 Fourth International Conference on Computational Intelligence and Communication Networks, Mathura, India, 3–5 November 2012 , 566 – 570 .

Nakawiro W , Erlich I. A combined GA-ANN strategy for solving optimal power flow with voltage security constraint . In: 2009 Asia-Pacific Power and Energy Engineering Conference, Wuhan, China, 27–31 March 2009 , 1 – 4 .

Sumpavakup C , Srikun I , Chusanapiputt S. A solution to the optimal power flow using artificial bee colony algorithm . In: 2010 International Conference on Power System Technology, Zhejiang, China, 24–28 October 2010 , 1 – 5 .

Abdellah D , Djamel L. Power flow analysis using adaptive neuro-fuzzy inference systems . In: 2015 3rd International Renewable and Sustainable Energy Conference (IRSEC), Marrakech, Morocco, 10–13 December 2015 , 1 – 5 .

Nemati M , Braun M , Tenbohlen S. Optimization of unit commitment and economic dispatch in microgrids based on genetic algorithm and mixed integer linear programming . Appl Energy , 2018 , 210 : 944 – 963 .

Alshareef A. An application of artificial intelligent optimization techniques to dynamic unit commitment for the western area of Saudi Arabia . In: 2011 Third International Conference on Computational Intelligence, Communication Systems and Networks, Bali, Indonesia, 26–28 July 2011 , 17 – 21 .

Arora I , Kaur M. Unit commitment scheduling by employing artificial neural network-based load forecasting . In: 2016 7th India International Conference on Power Electronics (IICPE), Patiala, India, 17–19 November 2016 , 1 – 6 .

Liu Z , Li N , Zhang C. Unit commitment scheduling using a hybrid ANN and Lagrangian relaxation method . In: 2008 International Conference on Multimedia and Ubiquitous Engineering (mue 2008), Busan, South Korea, 24–25 April 2008 , 481 – 484 .

Kumar VS , Mohan MR. Solution to security constrained unit commitment problem using genetic algorithm . International Journal of Electrical Power & Energy Systems , 2010 , 32 : 117 – 125 .

Prakash S , Sinha SK. Application of artificial intelligence in load frequency control of interconnected power system . International Journal of Engineering, Science and Technology , 2011 , 3 : 264 – 275 .

Moshtagh J , Rafinia A. A new approach to high impedance fault location in three-phase underground distribution system using combination of fuzzy logic & wavelet analysis . In: 2012 11th International Conference on Environment and Electrical Engineering, Venice, Italy, 15–18 May 2012 , 90 – 97 .

Santoso NT , Tan OT. Neural-net based realtime control of capacitors installed on distribution systems . IEEE Trans. Power Delivery , 1990 , 5 : 266 – 272 .

Ng HN , Salama MMA , Chikhani AY. Capacitor allocation by approximate reasoning: fuzzy capacitor placement . IEEE Trans Power Deliv , 2000 , 15 : 393 – 398 .

Zhang H , Zhang L , Meng F. Reactive power optimization based on genetic algorithm . In: POWERCON’98. International Conference on Power System Technology. Proceedings, Vol. 2 , Beijing, China, 18–21 August 1998 , 1448 – 1453 .

Mamandur KRC , Chenoweth RD. Optimal control of reactive power flow for improvements in voltage profiles and for real power loss minimization . IEEE Transactions on Power Apparatus and Systems, 1981 , PAS-100 : 3185 – 3194 .

Kothakotla V , Kumar B. Analysis and design of robust PID controller for grid voltage control of islanded microgrid using genetic algorithm . In: 2021 5th International Conference on Intelligent Computing and Control Systems (ICICCS), Madurai, India, 6–8 May 2021 , 719 – 726 .

Wang S , Duan J , Shi D , et al.  . A data-driven multi-agent autonomous voltage control framework using deep reinforcement learning . IEEE Trans Power Syst , 2020 , 35 : 4644 – 4654 .

Zidani Y , Zouggar S , Elbacha A. Steady-state analysis and voltage control of the self-excited induction generator using artificial neural network and an active filter . Iranian Journal of Science and Technology, Transactions of Electrical Engineering , 2018 , 42 : 41 – 48 .

Sumathi S. Artificial neural network application for voltage control and power flow control in power systems with UPFC . In: 2015 International Conference on Emerging Research in Electronics, Computer Science and Technology (ICERECT), Mandya, India, 17–19 December 2015 , 403 – 407 .

Kanata S , Sianipar GH , Maulidevi NU. Optimization of reactive power and voltage control in power system using hybrid artificial neural network and particle swarm optimization . In: 2018 2nd International Conference on Applied Electromagnetic Technology (AEMT), Lombok, Indonesia, 9–12 April 2018 , 67 – 72 .

Abdalla OH , Ghany AA , Fayek HH. Coordinated PID secondary voltage control of a power system based on genetic algorithm . In: 2016 Eighteenth International Middle East Power Systems Conference (MEPCON), Cairo, Egypt, 27–29 December 2016 , 214 – 219 .

Chung IY , Liu W , Cartes DA. Control parameter optimization for a microgrid system using particle swarm optimization . In: 2008 IEEE International Conference on Sustainable Energy Technologies, Singapore, 24–27 November 2008 , 837 – 842 .

Yousuf H , Zainal AY , Alshurideh M , et al.  . Artificial intelligence models in power system analysis . In: Hassanien AE , Bhatnagar R , Darwish A (eds). Artificial Intelligence for Sustainable Development: Theory, Practice and Future Applications . Cham, Switzerland : Springer , 2021 , 231 – 242 .

Aakula JL , Khanduri A , Sharma A. Determining reactive power levels to improve bus voltages using PSO . In: 2020 IEEE 17th India Council International Conference (INDICON), New Delhi, India, 10–13 December 2020 , 1 – 7 .

Karthikeyan R , Pasam S , Sudheer S , et al.  . Fuzzy fractional order PID based parallel cascade control system . In: Thampi S , Abraham A , Pal S , Rodriguez J (eds). Recent Advances in Intelligent Informatics. Advances in Intelligent Systems and Computing , Vol. 235 . Cham, Switzerland : Springer , 2014 , 293 – 302 .

Sallama A , Abbod M , Taylor G. Supervisory power system stability control using neuro-fuzzy system and particle swarm optimization algorithm . In: 2014 49th International Universities Power Engineering Conference (UPEC), Cluj-Napoca, Romania, 2–5 September 2014 , 1 – 6 .

Chen B , Li W , Yu P , et al.  . Planning approach for cross-regional optical transmission network supporting wide-area stability control services in power system . In: 2018 International Conference on Network Infrastructure and Digital Content (IC-NIDC), Guiyang, China, 22–24 August 2018 , 105 – 109 .

Torkzadeh R , NasrAzadani H , Aliabad AD , et al.  . A genetic algorithm optimized fuzzy logic controller for UPFC in order to damp of low frequency oscillations in power systems . In: 2014 22nd Iranian Conference on Electrical Engineering (ICEE), Tehran, Iran, 20–22 May 2014 , 706 – 712 .

Dutta S , Singh SP. Optimal rescheduling of generators for congestion management based on particle swarm optimization . IEEE Trans Power Syst , 2008 , 23 : 1560 – 1569 .

Nam S , Hur J. Probabilistic forecasting model of solar power outputs based on the naive Bayes classifier and kriging models . Energies , 2018 , 11 : 2982 .

Miraftabzadeh SM , Longo M , Foiadelli F , et al.  . Advances in the application of machine learning techniques for power system analytics: a survey . Energies , 2021 , 14 : 4776 .

Safari A , Babaei F , Farrokhifar M. A load frequency control using a PSO-based ANN for micro-grids in the presence of electric vehicles . Int J Ambient Energy , 2021 , 42 : 688 – 700 .

Joshi M , Sharma G , Davidson IE. Load frequency control of hydro electric system using application of fuzzy with particle swarm optimization algorithm . In: 2020 International Conference on Artificial Intelligence, Big Data, Computing and Data Communication Systems (icABCD), Durban, South Africa, 6–7 August 2020 , 1 – 6 .

Nguyen GN , Jagatheesan K , Ashour AS , et al.  . Ant colony optimization-based load frequency control of multi-area interconnected thermal power system with governor dead-band nonlinearity . In: Yang XS , Nagar AK , Joshi A (eds). Smart Trends in Systems, Security and Sustainability . Singapore : Springer , 2018 , 157 – 167 .

Balamurugan CR. Three area power system load frequency control using fuzzy logic controller . International Journal of Applied Power Engineering (IJAPE) , 2018 , 7 : 18 – 26 .

Otani T , Tanabe R , Koyanagi Y , et al.  . Cooperative load frequency control of generator and battery using a recurrent neural network . In: TENCON 2017–2017 IEEE Region 10 Conference, Penang, Malaysia, 5–8 November 2017 , 918 – 923 .

Kumar D , Mathur HD , Bhanot S , et al.  . Forecasting of solar and wind power using LSTM RNN for load frequency control in isolated microgrid . Int J Modelling Simul , 2021 , 41 : 311 – 323 .

Arora K , Kumar A , Kamboj VK , et al.  . Optimization methodologies and testing on standard benchmark functions of load frequency control for interconnected multi area power system in smart grids . Mathematics , 2020 , 8 : 980 .

Vadivelu KR , Marutheswar GV. Artificial intelligence technique based reactive power planning incorporating FACTS controllers in real time power transmission system . In: 2014 IEEE 2nd International Conference on Electrical Energy Systems (ICEES), Chennai, India, 7–9 January 2014 , 26 – 31 .

Khan MA , Algarni F. A healthcare monitoring system for the diagnosis of heart disease in the IoMT cloud environment using MSSO-ANFIS . IEEE Access , 2020 , 8 : 122259 – 122269 .

Liu M , Dong M , Wu C. A new ANFIS for parameter prediction with numeric and categorical inputs . IEEE Trans Autom Sci Eng , 2010 , 7 : 645 – 653 .

Ghadiri A , Haghifam MR , Larimi SMM. Comprehensive approach for hybrid AC/DC distribution network planning using genetic algorithm . IET Generation, Transmission & Distribution , 2017 , 11 : 3892 – 3902 .

Jain SK , Bhargava A , Pal RK. Three area power system load frequency control using fuzzy logic controller . In: 2015 International Conference on Computer, Communication and Control (IC4), Indore, India, 10–12 September 2015 , 1 – 6 .

Son YS , Kim HJ , Kim JT. A video-quality control scheme using ANFIS architecture in a DASH environment . Journal of Broadcast Engineering , 2018 , 23 : 104 – 114 .

Kannadasan K , Edla DR , Yadav MH , et al.  . Intelligent-ANFIS model for predicting measurement of surface roughness and geometric tolerances in three-axis CNC milling . IEEE Trans Instrum Meas , 2020 , 69 : 7683 – 7694 .

Penghui L , Ewees AA , Beyaztas BH , et al.  . Metaheuristic optimization algorithms hybridized with artificial intelligence model for soil temperature prediction: novel model . IEEE Access , 2020 , 8 : 51884 – 51904 .

Tawfeek TS , Ahmed AH , Hasan S. Analytical and particle swarm optimization algorithms for optimal allocation of four different distributed generation types in radial distribution networks . Energy Procedia , 2018 , 153 : 86 – 94 .

Yeom CU , Kwak KC. Adaptive neuro-fuzzy inference system predictor with an incremental tree structure based on a context-based fuzzy clustering approach . Applied Sciences , 2020 , 10 : 8495 .

Krasopoulos CT , Beniakar ME , Kladas AG. Multicriteria PM motor design based on ANFIS evaluation of EV driving cycle efficiency . IEEE Trans Transp Electrif , 2018 , 42 : 525 – 535 .

El-Hasnony IM , Barakat SI , Mostafa . Optimized ANFIS model using hybrid metaheuristic algorithms for Parkinson’s disease prediction in IoT environment . IEEE Access , 2020 , 8 : 119252 – 119270 .

Morshedizadeh M , Kordestani M , Carriveau R , et al.  . Power production prediction of wind turbines using a fusion of MLP and ANFIS networks . IET Renew Power Gener , 2018 , 12 : 1025 – 1033 .

Khosravi A , Nahavandi S , Creighton D. Prediction interval construction and optimization for adaptive neurofuzzy inference systems . IEEE Trans Fuzzy Syst , 2011 , 19 : 983 – 988 .

Elbaz K , Shen SL , Sun WJ , et al.  . Prediction model of shield performance during tunneling via incorporating improved particle swarm optimization into ANFIS . IEEE Access , 2020 , 8 : 39659 – 39671 .

Dovžan D , Škrjanc I. Fuzzy space partitioning based on hyperplanes defined by eigenvectors for Takagi-Sugeno fuzzy model identification . IEEE Trans Ind Electron , 2019 , 67 : 5144 – 5153 .

Castiello C , Fanelli AM , Lucarelli M , et al.  . Interpretable fuzzy partitioning of classified data with variable granularity . Appl Soft Comput , 2019 , 74 : 567 – 582 .

Alexandridis A , Chondrodima E , Sarimveis H. Radial basis function network training using a nonsymmetric partition of the input space and particle swarm optimization . IEEE Trans Neural Networks Learn Syst , 2012 , 24 : 219 – 230 .

Verstraete J. The spatial disaggregation problem: simulating reasoning using a fuzzy inference system . IEEE Trans Fuzzy Syst , 2016 , 25 : 627 – 641 .

Olamaei J , Moradi M , Kaboodi T. A new adaptive modified firefly algorithm to solve optimal capacitor placement problem . In: 18th Electric Power Distribution Conference, Kermanshah, Iran, 30 April–1 May 2013 , 1 – 6 .

Lee JS , Teng CL. An enhanced hierarchical clustering approach for mobile sensor networks using fuzzy inference systems . IEEE Internet Things J , 2017 , 4 : 1095 – 1103 .

Su ZG , Denoeux T. BPEC. Belief-peaks evidential clustering . IEEE Trans Fuzzy Syst , 2018 , 27 : 111 – 123 .

Xu P , Deng Z , Cui C , et al.  . Concise fuzzy system modeling integrating soft subspace clustering and sparse learning . IEEE Trans Fuzzy Syst , 2019 , 27 : 2176 – 2189 .

Sujil A , Kumar R , Bansal RC. FCM clustering-ANFIS-based PV and wind generation forecasting agent for energy management in a smart microgrid . The Journal of Engineering , 2019 , 2019 : 4852 – 4857 .

Neamatollahi P , Naghibzadeh M , Abrishami S. Fuzzy-based clustering-task scheduling for lifetime enhancement in wireless sensor networks . IEEE Sens J , 2017 , 17 : 6837 – 6844 .

Wang Z , Xu Y. A golden section-based double population genetic algorithm applied to reactive power optimization . IOP Conference Series: Earth and Environmental Science, 2021 , 645 : 012074 .

Kahouli O , Alsaif H , Bouteraa Y , et al.  . Power system reconfiguration in distribution network for improving reliability using genetic algorithm and particle swarm optimization . Applied Sciences , 2021 , 11 : 3092 .

Žarković SD , Stanković S , Shayesteh E , et al.  . Reliability improvement of distribution system through distribution system planning: MILP vs. GA . In: 2019 IEEE Milan PowerTech, Milan, Italy, 23–27 June 2019, 1 – 6 .

Ahmetovic H , Saric M , Hivziefendic J. Reliability based power distribution network planning using fuzzy logic . Advances in Electrical and Electronic Engineering , 2021 , 19 : 123 – 133 .

Suresh M , Sirish TS , Subhashini TV , et al.  . Load flow analysis of distribution system using artificial neural networks . In: Satapathy SC , Bhateja V , Udgata SK , Pattnaik PK (eds). Proceedings of the 5th International Conference on Frontiers in Intelligent Computing: Theory and Applications FICTA 2016, Vol. 1 . Springer , Singapore , 2017 , 515 – 524 .

Kumari M , Ranjan R , Singh VR. Optimal power distribution planning using improved particle swarm optimization . International Journal of Intelligent Systems and Applications in Engineering , 2018 , 3 : 170 – 177 .

Saha S , Mukherjee V. A novel multi-objective modified symbiotic organisms search algorithm for optimal allocation of distributed generation in radial distribution system . Neural Comput Appl , 2021 , 33 : 1751 – 1771 .

Hosseini MM , Parvania M. Artificial intelligence for resilience enhancement of power distribution systems . The Electricity Journal , 2021 , 34 : 106880 .

Lytras MD , Chui KT. The recent development of artificial intelligence for smart and sustainable energy systems and applications . Energies , 2019 , 12 : 3108 .

Gandhar S , Ohri J , Singh M. Dynamic reactive power optimization of hybrid micro grid in islanded mode using fuzzy tuned UPFC . J Inf Optim Sci , 2020 , 41 : 305 – 315 .

Tang Y , Hu W , Cao D , et al.  . Artificial intelligence-aided minimum reactive power control for the DAB converter based on harmonic analysis method . IEEE Trans Power Electron , 2021 , 36 : 9704 – 9710 .

Harrye YA , Ahmed KH , Aboushady AA. Reactive power minimization of dual active bridge DC/DC converter with triple phase shift control using neural network . In: 2014 International Conference on Renewable Energy Research and Application (ICRERA), Milwaukee, WI, USA, 19–22 October 2014 , 566 – 571 .

Wang Y , Li Y. A hybrid ant colony optimization algorithm for dynamic optimization of vehicle routing problem with time windows . Applied Intelligence , 2021 , 51 : 476 – 491 .

Sharma NK , Babu DS , Choube SC. Application of particle swarm optimization technique for reactive power optimization . In: IEEE International Conference on Advances in Engineering, Science and Management (ICAESM-2012), Nagapattinam, India, 30–31 March 2020 , 88 – 93 .

Bhattacharyya B , Gupta VK. Fuzzy based evolutionary algorithm for reactive power optimization with FACTS devices . International Journal of Electrical Power & Energy Systems , 2014 , 61 : 39 – 47 .

Bharti D. Multi-point optimal placement of shunt capacitor in radial distribution network: a comparison . In: 2020 International Conference on Emerging Frontiers in Electrical and Electronic Technologies (ICEFEET), Patna, India, 10–11 July 2020 , 1 – 6 .

Roy PK , Sultana S. Optimal reconfiguration of capacitor based radial distribution system using chaotic quasi oppositional chemical reaction optimization . Microsyst Technol , 2022 , 28 : 499 – 511 .

Pimentel Filho MC , De Lacerda EGM , Junior MM. Capacitor placement using ant colony optimization and gradient . In: 2009 15th International Conference on Intelligent System Applications to Power Systems, Curitiba, Brazil, 8–12 November 2009 , 1 – 4 .

Isac SJ , Kumar KS , Kumar PV. Optimal capacitor placement in radial distribution system to minimize the loss using fuzzy logic control . In: International Conference on Smart Structures and Systems—ICSSS; 13 , Chennai, India, 28–29 March 2013 , 33 – 40 .

Reddy MD , Reddy VV. Capacitor placement using fuzzy and particle swarm optimization method for maximum annual savings . ARPN Journal of Engineering and Applied Sciences , 2008 , 3 : 25 – 30 .

Shwehdi MH , Mohamed SR , Devaraj D. Optimal capacitor placement on West–East inter-tie in Saudi Arabia using genetic algorithm . Computers & Electrical Engineering , 2018 , 68 : 156 – 169 .

Mahdavian M , Kafi MH , Movahedi A , et al.  . Improve performance in electrical power distribution system by optimal capacitor placement using genetic algorithm . In: 2017 14th International Conference on Electrical Engineering/Electronics, Computer, Telecommunications, and Information Technology (ECTI-CON), Phuket, Thailand, 27–30 June 2017 , 749 – 752 .

Email alerts

Citing articles via.

• Advertising and Corporate Services

Affiliations

• Online ISSN 2515-396X
• Print ISSN 2515-4230
• Copyright © 2024 National Institute of Clean-and-Low-Carbon Energy
• About Oxford Academic
• Publish journals with us
• University press partners
• What we publish
• New features  
• Open access
• Institutional account management
• Rights and permissions
• Get help with access
• Accessibility
• Advertising
• Media enquiries
• Oxford University Press
• Oxford Languages
• University of Oxford

Oxford University Press is a department of the University of Oxford. It furthers the University's objective of excellence in research, scholarship, and education by publishing worldwide

• Copyright © 2024 Oxford University Press
• Cookie settings
• Cookie policy
• Privacy policy
• Legal notice

This Feature Is Available To Subscribers Only

Sign In or Create an Account

This PDF is available to Subscribers Only

For full access to this pdf, sign in to an existing account, or purchase an annual subscription.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Review Article
• Published: 09 February 2024

Artificial intelligence-based methods for renewable power system operation

• Yuanzheng Li   ORCID: orcid.org/0000-0001-8052-1233 1 , 2 ,
• Yizhou Ding   ORCID: orcid.org/0000-0003-3121-1487 3 ,
• Shangyang He   ORCID: orcid.org/0009-0000-2380-8617 3 ,
• Fei Hu   ORCID: orcid.org/0000-0003-2386-9035 1 ,
• Juntao Duan   ORCID: orcid.org/0000-0003-3957-7248 1 ,
• Guanghui Wen   ORCID: orcid.org/0000-0003-0070-8597 4 ,
• Hua Geng   ORCID: orcid.org/0000-0002-8336-6731 5 ,
• Zhengguang Wu 6 ,
• Hoay Beng Gooi   ORCID: orcid.org/0000-0002-5983-2181 7 ,
• Yong Zhao   ORCID: orcid.org/0009-0001-0977-6580 1 , 2 ,
• Chenghui Zhang   ORCID: orcid.org/0000-0003-2317-5930 8 ,
• Shengwei Mei   ORCID: orcid.org/0000-0002-2757-5977 9 &
• Zhigang Zeng   ORCID: orcid.org/0000-0003-4587-3588 1 , 2  

Nature Reviews Electrical Engineering volume  1 ,  pages 163–179 ( 2024 ) Cite this article

586 Accesses

2 Citations

2 Altmetric

Metrics details

• Energy grids and networks
• Energy management
• Renewable energy

Carbon neutrality goals are driving the increased use of renewable energy (RE). Large-scale use of RE requires accurate energy generation forecasts; optimized power dispatch, which minimizes costs while satisfying operational constraints; effective system control to ensure a stable power supply; and electricity markets that support bidding and trading decisions associated with RE. However, the uncertainties in RE generation make renewable power systems challenging to operate. For example, the intermittent nature of wind power can make it difficult to balance the supply and demand of electricity in real time; therefore, traditional power sources could be needed to meet the demand, which can increase electricity prices. This Review outlines the potential of artificial intelligence-based methods for supporting renewable power system operation. We discuss the ability of machine learning, deep learning and reinforcement learning methods to facilitate power system forecasts, dispatch, control and markets to support the use of RE. We also emphasize the applicability of these techniques to different operational problems. Finally, we discuss potential trends in renewable power system development and approaches to address the associated operational challenges such as the increasingly distributed nature of RE installations, diversification of energy storage systems and growing market complexity.

The large variabilities in renewable energy (RE) generation can make it challenging for renewable power systems to provide stable power supplies; however, artificial intelligence (AI)-based methods can help overcome these challenges.

Deep learning methods can provide accurate RE generation forecasts to help balance the supply of and demand for electricity.

Reinforcement learning techniques can effectively handle the increased computational complexity associated with optimizing power dispatch for renewable power systems to ensure that costs are minimized and operational constraints are met.

Renewable power systems are subject to greater instabilities than traditional systems, which can lead to voltage and frequency fluctuations in the power supply. AI-based techniques can provide real-time control signals to facilitate generation-to-demand control.

Reinforcement learning techniques can also be used to analyse market behaviours and optimize decision-making to support the effective integration of RE into power markets.

Future AI-based methods will need to solve the challenges that could arise from increases in the number of entities supplying RE and the diversity of energy storage systems, which will further complicate renewable power systems.

This is a preview of subscription content, access via your institution

Access options

Subscribe to this journal

Receive 12 digital issues and online access to articles

111,21 € per year

only 9,27 € per issue

Buy this article

• Purchase on Springer Link
• Instant access to full article PDF

Prices may be subject to local taxes which are calculated during checkout

Similar content being viewed by others

Data-driven assisted real-time optimal control strategy of submerged arc furnace via intelligent energy terminals considering large-scale renewable energy utilization

Deep learning based optimal energy management for photovoltaic and battery energy storage integrated home micro-grid system

A systematic review of the costs and impacts of integrating variable renewables into power grids

BP p.l.c. Energy outlook. BP https://www.bp.com/en/global/corporate/energy-economics/energy-outlook.html (2023).

Li, Y. et al. Deep reinforcement learning for smart grid operations: algorithms, applications, and prospects. Proc. IEEE 111 , 1055–1096 (2023).

Google Scholar  

Sharma, S., Xu, Y., Verma, A. & Panigrahi, B. K. Time-coordinated multi-energy management of smart buildings under uncertainties. IEEE Trans. Ind. Inform. 15 , 4788–4798 (2019).

China Monitoring and Early Warning Centre for Renewable Energy Absorption. Evaluation and analysis of Chinese renewable energy utilization in the fourth quarters of 2021. China Power http://www.chinapower.com.cn/zx/hyfx/20220315/138719.html (2022).

US Energy lnformation Administration. As Texas wind and solar capacity increase, energy curtailments are also likely to rise. EIA https://www.eia.gov/todayinenergy/detail.php?id=57100 (2023).

Li, Y., Zhao, Y., Wu, L. & Zeng, Z. Artificial Intelligence Enabled Computational Methods for Smart Grid Forecast and Dispatch (Springer Nature, 2023).

Razavi, S. & Tolson, B. A. A new formulation for feedforward neural networks. IEEE Trans. Neural Netw. 22 , 1588–1598 (2011).

PubMed   Google Scholar  

Ren, Z. et al. Embedding physics domain knowledge into a Bayesian network enables layer-by-layer process innovation for photovoltaics. NPJ Comput. Mater. 6 , 9 (2020).

ADS   CAS   Google Scholar  

Kingsford, C. & Salzberg, S. L. What are decision trees? Nat. Biotechnol. 26 , 1011–1013 (2008).

CAS   PubMed   PubMed Central   Google Scholar  

Noble, W. S. What is a support vector machine? Nat. Biotechnol. 24 , 1565–1567 (2006).

CAS   PubMed   Google Scholar  

Samek, W., Montavon, G., Lapuschkin, S., Anders, C. J. & Müller, K.-R. Explaining deep neural networks and beyond: a review of methods and applications. Proc. IEEE 109 , 247–278 (2021).

Parlos, A. G., Chong, K. T. & Atiya, A. F. Application of the recurrent multilayer perceptron in modeling complex process dynamics. IEEE Trans. Neural Netw. 5 , 255–266 (1994).

Krizhevsky, A., Sutskever, I. & Hinton, G. E. Imagenet classification with deep convolutional neural networks. Commun. ACM 60 , 84–90 (2017).

He, K., Zhang, X., Ren, S. & Sun, J. in Proc. IEEE Conf. Computer Vision and Pattern Recognition 770–778 (IEEE, 2016).

Pascanu, R., Mikolov, T. & Bengio, Y. in Int. Conf. Machine Learning 1310–1318 (2013).

Li, C. et al. Long short-term memory networks in memristor crossbar arrays. Nat. Mach. Intell. 1 , 49–57 (2019).

Ravanelli, M., Brakel, P., Omologo, M. & Bengio, Y. Light gated recurrent units for speech recognition. IEEE Trans. Emerg. Top. Comput. Intell. 2 , 92–102 (2018).

Bellemare, M. G., Dabney, W. & Munos, R. in Int. Conf. Machine Learning 449–458 (ACM, 2017).

Henderson, P. et al. in Proc. AAAI Conf. Artificial Intelligence Vol. 32 (AAAI, 2018).

Li, Y. et al. Dense skip attention based deep learning for day-ahead electricity price forecasting. IEEE Trans. Power Syst. 38 , 4308–4327 (2023).

ADS   Google Scholar  

Hong, T. et al. Energy forecasting: a review and outlook. Res. Pap. Econ. 7 , 376–388 (2020).

Liu, H. & Zhang, Z. A bi-party engaged modeling framework for renewable power predictions with privacy-preserving. IEEE Trans. Power Syst. 38 , 5794–5805 (2022).

Gu, Y. & Green, T. C. Power system stability with a high penetration of inverter-based resources. Proc. IEEE 111 , 832–853 (2023).

Wang, Y. et al. Short-term load forecasting for industrial customers based on TCN-LightGBM. IEEE Trans. Power Syst. 36 , 1984–1997 (2021).

Li, C., Tang, G., Xue, X., Saeed, A. & Hu, X. Short-term wind speed interval prediction based on ensemble GRU model. IEEE Trans. Sustain. Energy 11 , 1370–1380 (2019).

Zheng, Z. & Zhang, Z. A stochastic recurrent encoder decoder network for multistep probabilistic wind power predictions. IEEE Trans. Neural Netw. Learn. Syst. https://doi.org/10.1109/TNNLS.2023.3234130 (2023).

Article   PubMed   Google Scholar  

Chai, S., Xu, Z., Jia, Y. & Wong, W. K. A robust spatiotemporal forecasting framework for photovoltaic generation. IEEE Trans. Smart Grid 11 , 5370–5382 (2020).

Wang, J. et al. (eds.) Advances in Neural Information Processing Systems Vol. 35 5941–5954 (Curran Associates, 2022).

Wu, F., Jing, R., Zhang, X.-P., Wang, F. & Bao, Y. A combined method of improved grey BP neural network and MEEMD-ARIMA for day-ahead wave energy forecast. IEEE Trans. Sustain. Energy 12 , 2404–2412 (2021).

Dudek, G., Pełka, P. & Smyl, S. A hybrid residual dilated LSTM and exponential smoothing model for midterm electric load forecasting. IEEE Trans. Neural Netw. Learn. Syst. 33 , 2879–2891 (2021).

Dhiman, H. S., Deb, D. & Guerrero, J. M. Hybrid machine intelligent SVR variants for wind forecasting and ramp events. Renew. Sustain. Energy Rev. 108 , 369–379 (2019).

Tantithamthavorn, C., McIntosh, S., Hassan, A. E. & Matsumoto, K. The impact of automated parameter optimization on defect prediction models. IEEE Trans. Softw. Eng. 45 , 683–711 (2018).

Li, Z. et al. Deep learning based densely connected network for load forecasting. IEEE Trans. Power Syst. 36 , 2829–2840 (2021).

LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521 , 436–444 (2015).

ADS   CAS   PubMed   Google Scholar  

Yan, J. et al. Frequency-domain decomposition and deep learning based solar PV power ultra-short-term forecasting model. IEEE Trans. Ind. Appl. 57 , 3282–3295 (2021).

Zhou, X., Pang, C., Zeng, X., Jiang, L. & Chen, Y. A short-term power prediction method based on temporal convolutional network in virtual power plant photovoltaic system. IEEE Trans. Instrum. Meas. 72 , 1–10 (2023).

Abdel-Nasser, M., Mahmoud, K. & Lehtonen, M. Reliable solar irradiance forecasting approach based on choquet integral and deep LSTMs. IEEE Trans. Ind. Inform. 17 , 1873–1881 (2020).

Zheng, Z., Yang, L. & Zhang, Z. Conditional variational autoencoder informed probabilistic wind power curve modeling. IEEE Trans. Sustain. Energy 14 , 2445–2460 (2023).

Ziyabari, S., Du, L. & Biswas, S. K. Multibranch attentive gated ResNet for short-term spatio-temporal solar irradiance forecasting. IEEE Trans. Ind. Appl. 58 , 28–38 (2021).

Sun, C., Shrivastava, A., Singh, S. & Gupta, A. in Proc. IEEE Int. Conf. Computer Vision 843-852 https://doi.org/10.1109/ICCV.2017.97 (IEEE, 2017).

Zhou, H. et al. Informer: beyond efficient Transformer for long sequence time-series forecasting. Proc. AAAI Conf. Artif. Intell. 35 , 11106–11115 (2021).

Wang, C. et al. A transformer-based method of multi-energy load forecasting in integrated energy system. IEEE Trans. Smart Grid 13 , 2703–2714 (2022).

Verde, S. F. & Rossetto, N. The Future of Renewable Energy Communities in the EU: An Investigation at the Time of the Clean Energy Package (European Univ. Institute, 2020).

Sharda, S., Singh, M. & Sharma, K. RSAM: robust self-attention based multi-horizon model for solar irradiance forecasting. IEEE Trans. Sustain. Energy 12 , 1394–1405 (2020).

Li, Y., He, S., Li, Y., Ding, Q. & Zeng, Z. Renewable energy absorption oriented many-objective probabilistic optimal power flow. IEEE Trans. Netw. Sci. Eng . https://doi.org/10.1109/TNSE.2023.3290147 (2023).

Dommel, H. W. & Tinney, W. F. Optimal power flow solutions. IEEE Trans. Power Appar. Syst. PAS-87 , 1866–1876 (1968).

Frangioni, A., Gentile, C. & Lacalandra, F. Tighter approximated MILP formulations for unit commitment problems. IEEE Trans. Power Syst. 24 , 105–113 (2009).

Hou, H. et al. Data-driven economic dispatch for islanded micro-grid considering uncertainty and demand response. Int. J. Electr. Power Energy Syst. 136 , 107623 (2022).

Du, Y., Li, F., Li, J. & Zheng, T. Achieving 100× acceleration for n  – 1 contingency screening with uncertain scenarios using deep convolutional neural network. IEEE Trans. Power Syst. 34 , 3303–3305 (2019).

Zhou, M., Chen, M. & Low, S. H. DeepOPF-FT: one deep neural network for multiple AC-OPF problems with flexible topology. IEEE Trans. Power Syst. 38 , 964–967 (2023).

Li, Y., Wan, C., Chen, D. & Song, Y. Nonparametric probabilistic optimal power flow. IEEE Trans. Power Syst. 37 , 2758–2770 (2022).

Li, Y. et al. Optimal operation of multimicrogrids via cooperative energy and reserve scheduling. IEEE Trans. Ind. Inform. 14 , 3459–3468 (2018).

Owerko, D., Gama, F. & Ribeiro, A. in ICASSP 2020 — 2020 IEEE Int. Conf. Acoustics, Speech and Signal Processing 5930–5934 (IEEE, 2020).

Diehl, F. in 33rd Conf. Neural Information Processing Systems (NeurIPS 2019) 1–6 (MIT Press, 2019).

Yan, Z. & Xu, Y. Real-time optimal power flow: a lagrangian based deep reinforcement learning approach. IEEE Trans. Power Syst. 35 , 3270–3273 (2020).

Liu, S. et al. Varying condition SCOPF based on deep learning and knowledge graph. IEEE Trans. Power Syst. 38 , 3189–3200 (2023).

Singh, M. K., Kekatos, V. & Giannakis, G. B. Learning to solve the AC-OPF using sensitivity-informed deep neural networks. IEEE Trans. Power Syst. 37 , 2833–2846 (2022).

Liu, T. et al. A Bayesian learning based scheme for online dynamic security assessment and preventive control. IEEE Trans. Power Syst. 35 , 4088–4099 (2020).

Velloso, A. & Van Hentenryck, P. Combining deep learning and optimization for preventive security-constrained DC optimal power flow. IEEE Trans. Power Syst. 36 , 3618–3628 (2021).

Donti, P., Agarwal, A., Bedmutha, N. V., Pileggi, L. & Kolter, J. Z. Adversarially robust learning for security-constrained optimal power flow. Adv. Neural Inf. Process. Syst. 34 , 28677–28689 (2021).

Zeng, L. et al. Physics-constrained vulnerability assessment of deep reinforcement learning-based SCOPF. IEEE Trans. Power Syst. 38 , 2690–2704 (2023).

Yang, H., Yi, J., Zhao, J. & Dong, Z. Extreme learning machine based genetic algorithm and its application in power system economic dispatch. Neurocomputing 102 , 154–162 (2013).

Chen, W., Park, S., Tanneau, M. & Van Hentenryck, P. Learning optimization proxies for large-scale security-constrained economic dispatch. Electr. Power Syst. Res. 213 , 108566 (2022).

Han, J., Yan, L. & Li, Z. A task-based day-ahead load forecasting model for stochastic economic dispatch. IEEE Trans. Power Syst. 36 , 5294–5304 (2021).

Zhou, S. et al. Combined heat and power system intelligent economic dispatch: a deep reinforcement learning approach. Int. J. Electr. Energy Syst. 120 , 106016 (2020).

Yu, T., Zhou, B., Chan, K. W., Chen, L. & Yang, B. Stochastic optimal relaxed automatic generation control in non-Markov environment based on multi-step Q ( λ ) learning. IEEE Trans. Power Syst. 26 , 1272–1282 (2011).

Han, X., He, H., Wu, J., Peng, J. & Li, Y. Energy management based on reinforcement learning with double deep Q-learning for a hybrid electric tracked vehicle. Appl. Energy 254 , 113708 (2019).

CAS   Google Scholar  

Duan, J. et al. Deep-reinforcement-learning-based autonomous voltage control for power grid operations. IEEE Trans. Power Syst. 35 , 814–817 (2020).

Li, Y., He, S., Li, Y., Shi, Y. & Zeng, Z. Federated multiagent deep reinforcement learning approach via physics-informed reward for multi-microgrid energy management. IEEE Trans. Neural Netw. Learn. Syst . 1–13 https://doi.org/10.1109/TNNLS.2022.3232630 (2023).

Du, Y. & Li, F. Intelligent multi-microgrid energy management based on deep neural network and model-free reinforcement learning. IEEE Trans. Smart Grid 11 , 1066–1076 (2020).

Wang, Z., Liu, Y., Ma, Z., Liu, X. & Ma, J. LiPSG: lightweight privacy-preserving Q-learning-based energy management for the IoT-enabled smart grid. IEEE Internet Things J. 7 , 3935–3947 (2020).

Lee, S. & Choi, D.-H. Dynamic pricing and energy management for profit maximization in multiple smart electric vehicle charging stations: a privacy-preserving deep reinforcement learning approach. Appl. Energy 304 , 117754 (2021).

Chen, P., Liu, S., Chen, B. & Yu, L. Multi-agent reinforcement learning for decentralized resilient secondary control of energy storage systems against dos attacks. IEEE Trans. Smart Grid 13 , 1739–1750 (2022).

Liang, Y., Ding, Z., Zhao, T. & Lee, W.-J. Real-time operation management for battery swapping-charging system via multi-agent deep reinforcement learning. IEEE Trans. Smart Grid 14 , 559–571 (2022).

Salari, A., Ahmadi, S. E., Marzband, M. & Zeinali, M. Fuzzy Q-learning-based approach for real-time energy management of home microgrids using cooperative multi-agent system. Sustain. Cities Soc. 95 , 104528 (2023).

Padhy, N. Unit commitment—a bibliographical survey. IEEE Trans. Power Syst. 19 , 1196–1205 (2004).

Zhou, M., Wang, B. & Watada, J. Deep learning-based rolling horizon unit commitment under hybrid uncertainties. Energy 186 , 115843 (2019).

Ajagekar, A. & You, F. Deep reinforcement learning based unit commitment scheduling under load and wind power uncertainty. IEEE Trans. Sustain . Energy 14 , 803–812 (2023).

Li, F., Qin, J. & Zheng, W. X. Distributed Q-learning-based online optimization algorithm for unit commitment and dispatch in smart grid. IEEE Trans. Cybern. 50 , 4146–4156 (2020).

de Mars, P. & O’Sullivan, A. Applying reinforcement learning and tree search to the unit commitment problem. Appl. Energy 302 , 117519 (2021).

Dalal, G. & Mannor, S. Reinforcement learning for the unit commitment problem. 2015 IEEE Eindh. PowerTech . 1–6 https://doi.org/10.1109/PTC.2015.7232646 (2015).

Li, F., Jiahu, Q. & Wei, X. Z. Distributed Q-learning-based online optimization algorithm for unit commitment and dispatch in smart grid. IEEE Trans. Cybern. 50.9 , 4146–4156 (2019).

Ajagekar, A. & Fengqi, Y. Deep reinforcement learning based unit commitment scheduling under load and wind power uncertainty. IEEE Trans. Sustain. Energy 14.2 , 803–812 (2022).

Guedes, L. S., de Mendonça Maia, P., Lisboa, A. C., Vieira, D. A. G. & Saldanha, R. R. A unit commitment algorithm and a compact MILP model for short-term hydro-power generation scheduling. IEEE Trans. Power Syst. 32 , 3381–3390 (2016).

Shi, J. & Oren, S. S. Stochastic unit commitment with topology control recourse for power systems with large-scale renewable integration. IEEE Trans. Power Syst. 33 , 3315–3324 (2017).

Chen, Y. et al. Security-constrained unit commitment for electricity market: modeling, solution methods, and future challenges. IEEE Trans. Power Syst. 38 , 4668–4681 (2022).

Qu, K., Zheng, X. & Yu, T. Environmental–economic unit commitment with robust diffusion control of gas pollutants. IEEE Trans. Power Syst. 38 , 818–834 (2022).

Chu, Z. & Teng, F. Voltage stability constrained unit commitment in power systems with high penetration of inverter-based generators. IEEE Trans. Power Syst. 38 , 1572–1582 (2022).

Bertsimas, D., Litvinov, E., Sun, X. A., Zhao, J. & Zheng, T. Adaptive robust optimization for the security constrained unit commitment problem. IEEE Trans. Power Syst. 28 , 52–63 (2012).

Zhang, J. et al. A survey for solving mixed integer programming via machine learning. Neurocomputing 519 , 205–217 (2023).

Gasse, M., Chételat, D., Ferroni, N., Charlin, L. & Lodi, A. Exact combinatorial optimization with graph convolutional neural networks. Adv. Neural Information Processing Systems . 32, 1–13 (2019).

Paulus, M. B., Zarpellon, G., Krause, A., Charlin, L. & Maddison, C. in Int. Conf. Machine Learning 17584–17600 (PMLR, 2022).

Hatziargyriou, N. et al. Definition and classification of power system stability—revisited & extended. IEEE Trans. Power Syst. 36 , 3271–3281 (2021).

Parvez, I., Aghili, M., Sarwat, A. I., Rahman, S. & Alam, F. Online power quality disturbance detection by support vector machine in smart meter. J. Mod. Power Syst. Clean. Energy 7 , 1328–1339 (2019).

Impram, S., Varbak Nese, S. & Oral, B. Challenges of renewable energy penetration on power system flexibility: a survey. Energy Strateg. Rev. 31 , 100539 (2020).

Shi, Z. et al. Artificial intelligence techniques for stability analysis and control in smart grids: methodologies, applications, challenges and future directions. Appl. Energy 278 , 115733 (2020).

Shi, Z. et al. Bidirectional active transfer learning for adaptive power system stability assessment and dominant instability mode identification. IEEE Trans. Power Syst . 1–15 https://doi.org/10.1109/TPWRS.2022.3220955 (2022).

Xi, L., Wu, J., Xu, Y. & Sun, H. Automatic generation control based on multiple neural networks with actor–critic strategy. IEEE Trans. Neural Netw. Learn. Syst. 32 , 2483–2493 (2021).

Xi, L., Yu, L., Xu, Y., Wang, S. & Chen, X. A novel multi-agent DDQN-AD method-based distributed strategy for automatic generation control of integrated energy systems. IEEE Trans. Sustain. Energy 11 , 2417–2426 (2020).

Buşoniu, L., de Bruin, T., Tolić, D., Kober, J. & Palunko, I. Reinforcement learning for control: performance, stability, and deep approximators. Annu. Rev. Control. 46 , 8–28 (2018).

MathSciNet   Google Scholar  

Hwang, M., Muljadi, E., Jang, G. & Kang, Y. C. Disturbance-adaptive short-term frequency support of a DFIG associated with the variable gain based on the ROCOF and rotor speed. IEEE Trans. Power Syst. 32 , 1873–1881 (2017).

Kheshti, M. et al. Toward intelligent inertial frequency participation of wind farms for the grid frequency control. IEEE Trans. Ind. Inform. 16 , 6772–6786 (2020).

Liang, Y., Zhao, X. & Sun, L. A multiagent reinforcement learning approach for wind farm frequency control. IEEE Trans. Ind. Inform. 19 , 1725–1734 (2023).

Dong, H. & Zhao, X. Data-driven wind farm control via multiplayer deep reinforcement learning. IEEE Trans. Control. Syst. Technol. 31 , 1468–1475 (2023).

Yan, J., Nuertayi, A., Yan, Y., Liu, S. & Liu, Y. Hybrid physical and data driven modeling for dynamic operation characteristic simulation of wind turbine. Renew. Energy 215 , 118958 (2023).

Su, Y. et al. An adaptive PV frequency control strategy based on real-time inertia estimation. IEEE Trans. Smart Grid 12 , 2355–2364 (2021).

Chen, Y. & Xu, D. Review of soft-switching topologies for single-phase photovoltaic inverters. IEEE Trans. Power Electron. 37 , 1926–1944 (2022).

Zhao, B., Zhang, X. & Huang, J. AI algorithm-based two-stage optimal design methodology of high-efficiency CLLC resonant converters for the hybrid AC–DC microgrid applications. IEEE Trans. Ind. Electron. 66 , 9756–9767 (2019).

Cao, D. et al. Attention enabled multi-agent DRL for decentralized volt-VAR control of active distribution system using PV inverters and SVCs. IEEE Trans. Sustain. Energy 12 , 1582–1592 (2021).

Khan, M. A., Haque, A. & Kurukuru, V. S. B. Intelligent transition control approach for different operating modes of photovoltaic inverter. IEEE Trans. Ind. Appl. 58 , 2332–2340 (2022).

Alam, M. J. E., Muttaqi, K. M. & Sutanto, D. A multi-mode control strategy for VAr support by solar PV inverters in distribution networks. IEEE Trans. Power Syst. 30 , 1316–1326 (2015).

Weckx, S. & Driesen, J. Optimal local reactive power control by PV inverters. IEEE Trans. Sustain. Energy 7 , 1624–1633 (2016).

Fu, X., Li, S. & Jaithwa, I. Implement optimal vector control for LCL-filter-based grid-connected converters by using recurrent neural networks. IEEE Trans. Ind. Electron. 62 , 4443–4454 (2015).

Song, Y. et al. A Q-learning based robust MPC method for DFIG to suppress the rotor overcurrent. Int. J. Electr. Power Energy Syst. 141 , 108106 (2022).

Yin, Z., Wang, S. & Zhao, Q. Sequential reconfiguration of unbalanced distribution network with soft open points based on deep reinforcement learning. J. Mod. Power Syst. Clean. Energy 11 , 107–119 (2023).

Peng, F. Z. Flexible AC transmission systems (FACTS) and resilient AC distribution systems (RACDS) in smart grid. Proc. IEEE 105 , 2099–2115 (2017).

Zheng, D.-D., Madani, S. S. & Karimi, A. Data-driven distributed online learning control for islanded microgrids. IEEE J. Emerg. Sel. Top. Circuits Syst. 12 , 194–204 (2022).

Zamzam, A. S. & Sidiropoulos, N. D. Physics-aware neural networks for distribution system state estimation. IEEE Trans. Power Syst. 35 , 4347–4356 (2020).

Chen, Q., Lin, N., Bu, S., Wang, H. & Zhang, B. Interpretable time-adaptive transient stability assessment based on dual-stage attention mechanism. Power Syst. IEEE Trans. 38 , 2776–2790 (2023).

Ye, X., Yan, J., Wang, Y., Lu, L. & He, R. A novel capsule convolutional neural network with attention mechanism for high-voltage circuit breaker fault diagnosis. Electr. Power Syst. Res. 209 , 108003 (2022).

Wang, W., Yu, N., Gao, Y. & Shi, J. Safe off-policy deep reinforcement learning algorithm for Volt-VAR control in power distribution systems. IEEE Trans. Smart Grid 11 , 3008–3018 (2020).

Wang, Y., Mao, M., Chang, L. & Hatziargyriou, N. D. Intelligent voltage control method in active distribution networks based on averaged weighted double deep Q-network algorithm. J. Mod. Power Syst. Clean. Energy 11 , 132–143 (2023).

Sun, X. & Qiu, J. Two-stage Volt/Var control in active distribution networks with multi-agent deep reinforcement learning method. IEEE Trans. Smart Grid 12 , 2903–2912 (2021).

Cao, D. et al. Data-driven multi-agent deep reinforcement learning for distribution system decentralized voltage control with high penetration of PVs. IEEE Trans. Smart Grid 12 , 4137–4150 (2021).

Hu, D. et al. Multi-agent deep reinforcement learning for voltage control with coordinated active and reactive power optimization. IEEE Trans. Smart Grid 13 , 4873–4886 (2022).

Li, Y. et al. Many-objective distribution network reconfiguration via deep reinforcement learning assisted optimization algorithm. IEEE Trans. Power Deliv. 37 , 2230–2244 (2022).

Kushwaha, A., Gopal, M. & Singh, B. Q-learning based maximum power extraction for wind energy conversion system with variable wind speed. IEEE Trans. Energy Convers. 35 , 1160–1170 (2020).

Liu, H. & Wu, W. Federated reinforcement learning for decentralized voltage control in distribution networks. IEEE Trans. Smart Grid 13 , 3840–3843 (2022).

Poudyal, A. et al. Multiarea inertia estimation using convolutional neural networks and federated learning. IEEE Syst. J. 16 , 6401–6412 (2022).

Molina-García, A., Bouffard, F. & Kirschen, D. S. Decentralized demand-side contribution to primary frequency control. IEEE Trans. Power Syst. 26 , 411–419 (2011).

Wang, X., Wang, J. & Liu, J. Vehicle to grid frequency regulation capacity optimal scheduling for battery swapping station using deep Q-network. IEEE Trans. Ind. Inform. 17 , 1342–1351 (2021).

Li, S. et al. Battery protective electric vehicle charging management in renewable energy system. IEEE Trans. Ind. Inform. 19 , 1312–1321 (2023).

Kabir, M. E., Ghafouri, M., Moussa, B. & Assi, C. A two-stage protection method for detection and mitigation of coordinated EVSE switching attacks. IEEE Trans. Smart Grid 12 , 4377–4388 (2021).

Du, Y., Li, F., Kurte, K., Munk, J. & Zandi, H. Demonstration of intelligent HVAC load management with deep reinforcement learning: real-world experience of machine learning in demand control. IEEE Power Energy Mag. 20 , 42–53 (2022).

Wang, B., Li, Y., Ming, W. & Wang, S. Deep reinforcement learning method for demand response management of interruptible load. IEEE Trans. Smart Grid 11 , 3146–3155 (2020).

Hu, Z. et al. Intelligent and rapid event-based load shedding pre-determination for large-scale power systems: knowledge-enhanced parallel branching dueling Q-network approach. Appl. Energy 347 , 121468 (2023).

Zhang, B., Hu, W., Ghias, A. M. Y. M., Xu, X. & Chen, Z. Multi-agent deep reinforcement learning-based coordination control for grid-aware multi-buildings. Appl. Energy 328 , 120215 (2022).

Guelpa, E. & Verda, V. Demand response and other demand side management techniques for district heating: a review. Energy 219 , 119440 (2021).

Zhang, Z. et al. A review of technologies and applications on versatile energy storage systems. Renew. Sustain. Energy Rev. 148 , 111263 (2021).

Calero, F. et al. A review of modeling and applications of energy storage systems in power grids. Proc. IEEE. 111 , 806–831 (2023).

Yao, F., Zhao, J., Li, X., Mao, L. & Qu, K. RBF neural network based virtual synchronous generator control with improved frequency stability. IEEE Trans. Ind. Inform. 17 , 4014–4024 (2021).

Wang, Y. & Wai, R.-J. Adaptive fuzzy-neural-network power decoupling strategy for virtual synchronous generator in micro-grid. IEEE Trans. Power Electron. 37 , 3878–3891 (2022).

Saadatmand, S., Shamsi, P. & Ferdowsi, M. Adaptive critic design-based reinforcement learning approach in controlling virtual inertia-based grid-connected inverters. Int. J. Electr. Power Energy Syst. 127 , 106657 (2021).

Bui, V.-H., Hussain, A. & Kim, H.-M. Double deep Q -learning-based distributed operation of battery energy storage system considering uncertainties. IEEE Trans. Smart Grid 11 , 457–469 (2020).

Shuai, H., Li, F., Pulgar-Painemal, H. & Xue, Y. Branching dueling Q-network-based online scheduling of a microgrid with distributed energy storage systems. IEEE Trans. Smart Grid 12 , 5479–5482 (2021).

Hosseini, M. M. & Parvania, M. Hierarchical intelligent operation of energy storage systems in power distribution grids. IEEE Trans. Sustain. Energy 14 , 741–750 (2023).

Yao, Z. et al. Machine learning for a sustainable energy future. Nat. Rev. Mater. 8 , 202–215 (2022).

ADS   PubMed   PubMed Central   Google Scholar  

Ye, Y., Qiu, D., Sun, M., Papadaskalopoulos, D. & Strbac, G. Deep reinforcement learning for strategic bidding in electricity markets. IEEE Trans. Smart Grid 11 , 1343–1355 (2019).

Liang, Y., Guo, C., Ding, Z. & Hua, H. Agent-based modeling in electricity market using deep deterministic policy gradient algorithm. IEEE Trans. Power Syst. 35 , 4180–4192 (2020).

Zhang, Z., Chen, Z. & Lee, W.-J. Soft actor–critic algorithm featured residential demand response strategic bidding for load aggregators. IEEE Trans. Ind. Appl. 58 , 4298–4308 (2022).

Peng, F. et al. Review on bidding strategies for renewable energy power producers participating in electricity spot markets. Sustain. Energy Technol. Assess. 58 , 103329 (2023).

Longoria, G., Davy, A. & Shi, L. Subsidy-free renewable energy trading: a meta agent approach. IEEE Trans. Sustain. Energy 11 , 1707–1716 (2019).

Ochoa, T., Gil, E., Angulo, A. & Valle, C. Multi-agent deep reinforcement learning for efficient multi-timescale bidding of a hybrid power plant in day-ahead and real-time markets. Appl. Energy 317 , 119067 (2022).

Jeong, J., Kim, S. W. & Kim, H. Deep reinforcement learning based real-time renewable energy bidding with battery control. IEEE Trans. Energy Mark. Policy Regul. 1 , 85–96 (2023).

Tang, Q., Guo, H. & Chen, Q. Multi-market bidding behavior analysis of energy storage system based on inverse reinforcement learning. IEEE Trans. Power Syst. 37 , 4819–4831 (2022).

Lin, Z. & Liu, X. Wind power forecasting of an offshore wind turbine based on high-frequency SCADA data and deep learning neural network. Energy 201 , 117693 (2020).

Jin, X., Xu, Z. & Qiao, W. Condition monitoring of wind turbine generators using SCADA data analysis. IEEE Trans. Sustain. Energy 12 , 202–210 (2021).

Yu, Y., Chen, L. & Liu, R. The source of wind power producers’ market power. Energy Policy 173 , 113401 (2023).

Tellidou, A. C. & Bakirtzis, A. G. Agent-based analysis of capacity withholding and tacit collusion in electricity markets. IEEE Trans. Power Syst. 22 , 1735–1742 (2007).

Razmi, P., Buygi, M. O. & Esmalifalak, M. Collusion strategy investigation and detection for generation units in electricity market using supervised learning paradigm. IEEE Syst. J. 15 , 146–157 (2020).

Qiu, D. et al. Strategic retail pricing and demand bidding of retailers in electricity market: a data-driven chance-constrained programming. Adv. Appl. Energy 7 , 100100 (2022).

Tsaousoglou, G. et al. Flexibility aggregation of temporally coupled resources in real-time balancing markets using machine learning. IEEE Trans. Ind. Inform. 18 , 4342–4351 (2021).

Chen, T. & Su, W. Indirect customer-to-customer energy trading with reinforcement learning. IEEE Trans. Smart Grid 10 , 4338–4348 (2018).

Taghizadeh, A., Montazeri, M. & Kebriaei, H. Deep reinforcement learning-aided bidding strategies for transactive energy market. IEEE Syst. J. 16 , 4445–4453 (2022).

Qian, T., Shao, C., Shi, D., Wang, X. & Wang, X. Automatically improved VCG mechanism for local energy markets via deep learning. IEEE Trans. Smart Grid 13 , 1261–1272 (2021).

Zhao, Z., Feng, C. & Liu, A. L. Comparisons of auction designs through multiagent learning in peer-to-peer energy trading. IEEE Trans. Smart Grid 14 , 593–605 (2022).

Ye, Y., Papadaskalopoulos, D., Yuan, Q., Tang, Y. & Strbac, G. Multi-agent deep reinforcement learning for coordinated energy trading and flexibility services provision in local electricity markets. IEEE Trans. Smart Grid 14 , 1541–1554 (2023).

Wang, J., Mishra, D. K., Li, L. & Zhang, J. Demand side management and peer-to-peer energy trading for industrial users using two-level multi-agent reinforcement learning. IEEE Trans. Energy Mark. Policy Regul. 1 , 23–36 (2023).

Gao, G., Wen, Y. & Tao, D. Distributed energy trading and scheduling among microgrids via multiagent reinforcement learning. IEEE Trans. Neural Networks Learn. Syst . 1–15 https://doi.org/10.1109/TNNLS.2022.3170070 (2022).

Nunna, H. K., Sesetti, A., Rathore, A. K. & Doolla, S. Multiagent-based energy trading platform for energy storage systems in distribution systems with interconnected microgrids. IEEE Trans. Ind. Appl. 56 , 3207–3217 (2020).

Lu, X. et al. Reinforcement learning-based microgrid energy trading with a reduced power plant schedule. IEEE Internet Things J. 6 , 10728–10737 (2019).

Yan, L., Chen, X., Chen, Y. & Wen, J. A hierarchical deep reinforcement learning-based community energy trading scheme for a neighborhood of smart households. IEEE Trans. Smart Grid 13 , 4747–4758 (2022).

Wu, Y., Zhao, T., Yan, H., Liu, M. & Liu, N. Hierarchical hybrid multi-agent deep reinforcement learning for peer-to-peer energy trading among multiple heterogeneous microgrids. IEEE Trans. Smart Grid 14 , 4649–4665 (2023).

Zamee, M. A., Han, D. & Won, D. Online hour-ahead load forecasting using appropriate time-delay neural network based on multiple correlation–multicollinearity analysis in IoT energy network. IEEE Internet Things J. 9 , 12041–12055 (2022).

Zhang, X., Pipattanasomporn, M., Chen, T. & Rahman, S. An IoT-based thermal model learning framework for smart buildings. IEEE Internet Things J. 7 , 518–527 (2020).

Lin, W., Wu, D. & Boulet, B. Spatial-temporal residential short-term load forecasting via graph neural networks. IEEE Trans. Smart Grid 12 , 5373–5384 (2021).

Zheng, X. et al. A multi-scale time-series dataset with benchmark for machine learning in decarbonized energy grids. Sci. Data 9 , 359 (2022).

PubMed   PubMed Central   Google Scholar  

Yin, X., Zhu, Y. & Hu, J. A comprehensive survey of privacy-preserving federated learning: a taxonomy, review, and future directions. ACM Comput. Surv. 54 , 1–36 (2021).

Hu, Q., Guo, Z. & Li, F. Imitation learning based fast power system production cost minimization simulation. IEEE Trans. Power Syst. 38 , 2951–2954 (2023).

Bellemare, M. G., Dabney, W. & Rowland, M. Distributional Reinforcement Learning (MIT Press, 2023).

Hu, Q., Zhang, S., Yu, M. & Xie, Z. Short-term wind speed or power forecasting with heteroscedastic support vector regression. IEEE Trans. Sustain. Energy 7 , 241–249 (2016).

Download references

Acknowledgements

This work was supported in part by the National Key R&D Program of China (grant 2021ZD0201300), National Natural Science Foundation of China (grants 62325304 and 62073148) and Key Project of the National Natural Science Foundation of China (grant 62233006).

Author information

Authors and affiliations.

Key Laboratory of lmage Information Processing and Intelligent Control of Ministry of Education of China, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China

Yuanzheng Li, Fei Hu, Juntao Duan, Yong Zhao & Zhigang Zeng

National Key Laboratory of Multispectral Information Intelligent Processing Technology, Huazhong University of Science and Technology, Wuhan, China

Yuanzheng Li, Yong Zhao & Zhigang Zeng

China-EU Institute for Clean and Renewable Energy, Huazhong University of Science and Technology, Wuhan, China

Yizhou Ding & Shangyang He

Department of Systems Science, School of Mathematics, Southeast University, Nanjing, China

Guanghui Wen

Department of Automation, Tsinghua University, Beijing, China

State Key Laboratory of Industrial Control Technology, Institute of Cyber-Systems and Control, Zhejiang University, Hangzhou, China

Zhengguang Wu

School of Electrical and Electronics Engineering, Nanyang Technological University, Singapore, Singapore

Hoay Beng Gooi

School of Control Science and Engineering, Shandong University, Jinan, China

Chenghui Zhang

Department of Electrical Engineering, Tsinghua University, Beijing, China

Shengwei Mei

You can also search for this author in PubMed   Google Scholar

Contributions

Y.L., Y.D., S.H., F.H. and J.D. researched data for the article. All authors contributed substantially to discussion of the content. Y.L., Y.D., S.H., F.H., J.D., G.W. and H.B.G. wrote the article. All authors edited the manuscript before submission.

Corresponding authors

Correspondence to Guanghui Wen , Hoay Beng Gooi or Zhigang Zeng .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Peer review

Peer review information.

Nature Reviews Electrical Engineering thanks Bhuvanagiri Ramesh and the other, anonymous, reviewer for their contribution to the peer review of this work.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Cite this article.

Li, Y., Ding, Y., He, S. et al. Artificial intelligence-based methods for renewable power system operation. Nat Rev Electr Eng 1 , 163–179 (2024). https://doi.org/10.1038/s44287-024-00018-9

Download citation

Accepted : 20 December 2023

Published : 09 February 2024

Issue Date : March 2024

DOI : https://doi.org/10.1038/s44287-024-00018-9

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

• News & Media
• Chemical Biology
• Computational Biology
• Ecosystem Science
• Cancer Biology
• Exposure Science & Pathogen Biology
• Metabolic Inflammatory Diseases
• Advanced Metabolomics
• Mass Spectrometry-Based Measurement Technologies
• Spatial and Single-Cell Proteomics
• Structural Biology
• Biofuels & Bioproducts
• Human Microbiome
• Soil Microbiome
• Synthetic Biology
• Computational Chemistry
• Chemical Separations
• Chemical Physics
• Atmospheric Aerosols
• Human-Earth System Interactions
• Modeling Earth Systems
• Coastal Science
• Plant Science
• Subsurface Science
• Terrestrial Aquatics
• Materials in Extreme Environments
• Precision Materials by Design
• Science of Interfaces
• Friction Stir Welding & Processing
• Dark Matter
• Flavor Physics
• Fusion Energy Science
• Neutrino Physics
• Quantum Information Sciences
• Emergency Response
• AGM Program
• Tools and Capabilities
• Grid Architecture
• Grid Cybersecurity
• Grid Energy Storage
• Earth System Modeling
• Energy System Modeling
• Transmission
• Distribution
• Appliance and Equipment Standards
• Building Energy Codes
• Advanced Building Controls
• Advanced Lighting
• Building-Grid Integration
• Building and Grid Modeling
• Commercial Buildings
• Federal Performance Optimization
• Resilience and Security
• Grid Resilience and Decarbonization
• Building America Solution Center
• Energy Efficient Technology Integration
• Home Energy Score
• Electrochemical Energy Storage
• Flexible Loads and Generation
• Grid Integration, Controls, and Architecture
• Regulation, Policy, and Valuation
• Science Supporting Energy Storage
• Chemical Energy Storage
• Waste Processing
• Radiation Measurement
• Environmental Remediation
• Subsurface Energy Systems
• Carbon Capture
• Carbon Storage
• Carbon Utilization
• Advanced Hydrocarbon Conversion
• Fuel Cycle Research
• Advanced Reactors
• Reactor Operations
• Reactor Licensing
• Solar Energy
• Wind Resource Characterization
• Wildlife and Wind
• Community Values and Ocean Co-Use
• Wind Systems Integration
• Wind Data Management
• Distributed Wind
• Energy Equity & Health
• Environmental Monitoring for Marine Energy
• Marine Biofouling and Corrosion
• Marine Energy Resource Characterization
• Testing for Marine Energy
• The Blue Economy
• Environmental Performance of Hydropower
• Hydropower Cybersecurity and Digitalization
• Hydropower and the Electric Grid
• Materials Science for Hydropower
• Pumped Storage Hydropower
• Water + Hydropower Planning
• Grid Integration of Renewable Energy
• Geothermal Energy
• Algal Biofuels
• Aviation Biofuels
• Waste-to-Energy and Products
• Hydrogen & Fuel Cells
• Emission Control
• Energy-Efficient Mobility Systems
• Lightweight Materials
• Vehicle Electrification
• Vehicle Grid Integration
• Contraband Detection
• Pathogen Science & Detection
• Explosives Detection
• Threat-Agnostic Biodefense
• Discovery and Insight
• Proactive Defense
• Trusted Systems
• Nuclear Material Science
• Radiological & Nuclear Detection
• Nuclear Forensics
• Ultra-Sensitive Nuclear Measurements
• Nuclear Explosion Monitoring
• Global Nuclear & Radiological Security
• Disaster Recovery
• Global Collaborations
• Legislative and Regulatory Analysis
• Technical Training
• Additive Manufacturing
• Deployed Technologies
• Rapid Prototyping
• Systems Engineering
• 5G Security
• RF Signal Detection & Exploitation
• Climate Security
• Internet of Things
• Maritime Security
• Artificial Intelligence
• Graph and Data Analytics
• Software Engineering
• Computational Mathematics & Statistics
• High-Performance Computing
• Visual Analytics
• Lab Objectives
• Publications & Reports
• Featured Research
• Diversity, Equity, Inclusion & Accessibility
• Lab Leadership
• Lab Fellows
• Staff Accomplishments
• Undergraduate Students
• Graduate Students
• Post-graduate Students
• University Faculty
• University Partnerships
• K-12 Educators and Students
• STEM Workforce Development
• STEM Outreach
• Meet the Team
• Internships
• Regional Impact
• Philanthropy
• Volunteering
• Available Technologies
• Industry Partnerships
• Licensing & Technology Transfer
• Entrepreneurial Leave
• Atmospheric Radiation Measurement User Facility
• Electricity Infrastructure Operations Center
• Energy Sciences Center
• Environmental Molecular Sciences Laboratory
• Grid Storage Launchpad
• Institute for Integrated Catalysis
• Interdiction Technology and Integration Laboratory
• PNNL Portland Research Center
• PNNL Seattle Research Center
• PNNL-Sequim (Marine and Coastal Research)
• Radiochemical Processing Laboratory
• Shallow Underground Laboratory

Artificial Intelligence/Machine Learning Technology in Power System Applications

Published: April 26, 2024

Research topics

We apologize for the inconvenience...

To ensure we keep this website safe, please can you confirm you are a human by ticking the box below.

If you are unable to complete the above request please contact us using the below link, providing a screenshot of your experience.

https://ioppublishing.org/contacts/

Artificial intelligence (AI)-based optimization of power electronic converters for improved power system stability and performance

Ieee account.

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

Academia.edu no longer supports Internet Explorer.

To browse Academia.edu and the wider internet faster and more securely, please take a few seconds to  upgrade your browser .

Enter the email address you signed up with and we'll email you a reset link.

• We're Hiring!
• Help Center

Artificial Intelligence in Power Station

2019, International Journal for Research in Applied Science and Engineering Technology

Related Papers

IOP Conference Series: Earth and Environmental Science

firas alnaimi

Pritish Dey

The development of a health monitoring system for electrical generators using AI techniques is presented. It focuses on Integrating AI techniques with Fuzzy based monitoring system for Generators health monitoring. Two AI systems such as ANN and FLS are simulated and tested to evaluate the performance monitoring for generators. A comparison between the two AI methods are done. Fuzzy Logic System was found to have a higher efficiency over ANN. For the possible causes of the fault, ANN and FLS have produced 70% and 100% efficient respectively. Keywords— Generator health monitoring, ANN-based monitoring, Partial discharge, generator temperature, fuzzy logic.

JESUS FILANDER CARATAR CHAUX

Science and Technology of Nuclear Installations

Economic constraints are driving the electric power industry to seek improved methods for monitoring, control, and diagnostics. To increase plant availability, various techniques have been implemented in industry to assess equipment condition to prevent system inoperability. The availability of a large number of measured signals and additional component information and the increasing number of signal processing options to analyze sampled data motivate the assimilation of such diverse information into a plantwide condition monitor. The use of fuzzy logic is described herein for the purpose of performing the decision making regarding the system status and the possible need for component maintenance. Fuzzy-logic-based diagnostic monitoring is applied to data acquired from instrumentation within operating facilities.

International Journal for Research in Applied Science and Engineering Technology IJRASET

IJRASET Publication

In this paper, the fuzzy logic is employed to Fault Identification in grid Network by Using Rule-based fuzzy-logic. The reliability of power supply is mostly full of the occurrence of faults within the power system network. Recent advances within the field of fuzzy logic systems and variety of successful applications in electric power (wattage) systems show that logic is efficiently applied to house imprecision, ambiguity and probabilistic information in the input file to scale back the time of power outages and for the immediate restoration, the fault identification method must be very effective. In case of a fault, the system operator gets very difficult information to choose on restoring the tripped feeder in normal operation. Therefore, in order to effectively build the ability to make decisions, a fuzzy logic-based fault diagnosis method has been introduced in this research. This method can identify the types of faults that occur in the radial-topology of the power system network. In that method, three-phase feeder current, phase voltage, and neutral current are used as fuzzy-inference system (FIS) input. Keywords: Distribution system network, fault diagnosis, and Rule based fuzzy logic. I. INTRODUCTION Nowadays, the substantive increase in complexity and variety of the consuming market, the electrical power grid has been demanding considerable updating and also a big improvement within the monitoring, control, and protection equipment. On the distribution system, faulty space is usually estimated based on the knowledge provided by the customer and also the accumulated experience in the technical field, which helps guide the support team in suspicious places. The faults of the facility system affect the general performance of the facility system. The factor like lightning, equipment damage, tree, animal, natural disaster, and human affect the facility system operation. These factors aren't predictable. When outages occur, fast and proper restoration of the feeder is vital to take care of the great power quality and customer satisfaction. Hence, to scale back the time of outages and immediate restoration, the expert systems are needed to spot the faults of the facility system. Failure of insulation of distribution system, bridging of excited phase conductors by objects, accident e.t.c. These events affect the voltage and current value of the distribution system and sometimes the entire power system. Given the fact that bulk distribution systems are overhead and have radial topologies, the need for accurate and reliable fault detection systems becomes much easier. In recent times, researchers have become more interested in finding solutions to ambiguities, incomplete fault information, errors in fault data, and knowledge redundancies [1]. Faults are usually taken care of by the interconnection of protection devices like relays, Current transformers, Voltage Transformers, Circuit Breakers, and fuses that detect the fault and eventually isolate the affected sections from the whole power grid. Faults are often of varied types and their detection and clearing approach may differ [2]. Voltage & current sensors have installed on the transmission line for real-time implementation and their cost is high. Matlab 7.5 for finding and analyzing defects in facility distribution network. The software has developed an Artificial neural network (ANN) based program, designed to detect defects [3]. Power grids round the globe are undergoing a huge transformation towards smart power grids with the assistance of rapidly developing monitoring and control methodology. Among these detecting, the fault and therefore the phase which underwent the fault is of great importance. Classification of the fault has the world of interest for varied researchers and as an outcome; several fault classification methods are implemented over time. a number of the prominent methods are Neural network-based technique, wavelet transforms based technique, fuzzy and fuzzy-network based technique, etc[4]. it's generally independent of protective relaying. Protecting assets through short-term generation of visit signals is not the key to reducing the workload of a missed department as quickly as possible. The goal is to provide information to the grid operator about errors prior to on-site inspections. Transmission lines are most susceptible to the occurrence of the fault. Fault detection, direction estimation, and faulty phase selection play a critical role within the protection of a cable. Accurate and fast fault detection and classification under a spread of fault conditions are important requirements of any protective relaying scheme [5].

Anais do 11. Congresso Brasileiro de Inteligência Computacional

Elaine Inacio Bueno

subodh panda

In the power generation industry, any failures may lead to blackouts which can have very severe impact on many (or actually most) aspects of our life and can cripple many services, which are crucial to our well-being. Therefore, minimizing outages and preventing plant equipment failures is becoming a critical issue for power plant operators. While condition monitoring of The most critical machines as turbines, compressors, like ID fans, FD fans ,PA fans, BFP, CWP, CEP, Mill Motors , boiler feed pumps and other big machinery, seem to be receiving the same level of attention from engineers. Failure of these machines may have catastrophic impact on the power plant operation, therefore it was decided to investigate a possibility of developing a system, which would trigger a warning in case of suspicious machine behavior. The aim of the research work was to analyze data in order to develop an alarm system which signals a warning in case of abnormal state of operation and, consequently, c...

shanti swarup

Tawanda Mushiri

The final case study application was at a thermal power generation company. In large-scale industrial applications, the controlling and optimization of the parameters must be done efficiently and effectively so as to attain smooth operation of the plant. In this research, the main control parameters in the boiler were identified as coal flows, temperatures in the combustion zone, airto-fuel ratio and ash content, that is, the percentage of ash in raw coal, and mineral content. These parameters are monitored to avoid clinker formation in the super heater tubes of the boiler. Condition-based maintenance (CBM) approaches were used to monitor the boiler parameters. Fuzzy logic was applied in the monitoring of these parameters.

Information Sciences

Silvio Simani

RELATED PAPERS

Kate McCallin

A. Panfilov

Science Advances

Abdoulaye Fall

Analytica. Revista de Filosofia

Ricardo Terra

South African Journal of Philosophy

Carin Robinson

Child Indicators Research

Gemma Crous

Journal of microbiology, epidemiology and immunobiology

Inna Malchenko

ACS Medicinal Chemistry Letters

Michelle Palmer

elsa cariou

The Journal of Clinical Endocrinology &amp; Metabolism

Neil Chapman

ChemElectroChem

Sreekumar Kurungot

Annisa Jannah07

Somma: Revista Cientifica do Instituto Federal de Educação Ciência e Tecnologia do Piauí

Alysson Alves

International Journal of Renewable Energy Research

Annals of Agricultural and Environmental Medicine

Aleksandra Kasperczyk

哪里买新加坡国立大学毕业证 nus毕业证双学位证书教育部认证

Food and Nutrition Sciences

sadia sultana

2011 17th IEEE Real-Time and Embedded Technology and Applications Symposium

Rodolfo Pellizzoni

Aydın Gürel

European Journal of Medical and Health Sciences

RELATED TOPICS

•   We're Hiring!
•   Help Center
• Find new research papers in:
• Health Sciences
• Earth Sciences
• Cognitive Science
• Mathematics
• Computer Science
• Academia ©2024

International Journal of Trend in Scientific Research and Development

Best international journal, open access journal issn: 2456-6470 , impact factor: 8.125.

• About IJTSRD
• Editor Board
• Reviewer Member
• Conference Proposal
• Special Issue Proposal
• Join as Reviewer
• Current Issue
• Research Area
• Call For Paper
• Publication Guide
• Processing Charge
• Sample Paper Format
• Copyright Form
• Indexing & Abstracting
• Ethics Policy
• Special Issue
• Submit Paper

Artificial Intelligence in Power Station

Call for papers.

• Call for Paper
• Publication Guidelines
• Submit Paper Online
• Author Account

International Journal of Trend in Scientific Research and Development - IJTSRD having online ISSN 2456-6470. IJTSRD is a leading Open Access, Peer-Reviewed International Journal which provides rapid publication of your research articles and aims to promote the theory and practice along with knowledge sharing between researchers, developers, engineers, students, and practitioners working in and around the world in many areas like Sciences, Technology, Innovation, Engineering, Agriculture, Management and many more and it is recommended by all Universities, review articles and short communications in all subjects. IJTSRD running an International Journal who are proving quality publication of peer reviewed and refereed international journals from diverse fields that emphasizes new research, development and their applications. IJTSRD provides an online access to exchange your research work, technical notes & surveying results among professionals throughout the world in e-journals. IJTSRD is a fastest growing and dynamic professional organization. The aim of this organization is to provide access not only to world class research resources, but through its professionals aim to bring in a significant transformation in the real of open access journals and online publishing.

• Submit Manuscript
• Terms & Condition
• Disclaimer Policy
• Copyright Claims Policy
• Privacy Policy
• Shipping & Delivery Policy
• Cancellation & Refund Policy

Stay Connected

Sign in with your account, forgot password | register now.

RIOT-ML: toolkit for over-the-air secure updates and performance evaluation of TinyML models

• Open access
• Published: 22 May 2024

Cite this article

You have full access to this open access article

• Zhaolan Huang 1 ,
• Koen Zandberg 1 ,
• Kaspar Schleiser 1 &
• Emmanuel Baccelli 1 , 2  

Practitioners in the field of TinyML lack so far a comprehensive, “batteries-included” toolkit to streamline continuous integration, continuous deployment and performance assessments of executing diverse machine learning models on various low-power IoT hardware. Addressing this gap, our paper introduces RIOT-ML, a versatile toolkit crafted to assist IoT designers and researchers in these tasks. To this end, we designed RIOT-ML based on an integration of an array of functionalities from a low-power embedded OS, a universal model transpiler and compiler, a toolkit for TinyML performance measurement, and a low-power over-the-air secure update framework—all of which usable on an open-access IoT testbed available to the community. Our open-source implementation of RIOT-ML and the initial experiments we report on showcase its utility in experimentally evaluating TinyML model performance across fleets of low-power IoT boards under test in the field, featuring a wide spectrum of heterogeneous microcontroller architectures and fleet network connectivity configurations. The existence of an open-source toolkit such as RIOT-ML is essential to expedite research combining artificial intelligence and IoT and to foster the full realization of edge computing’s potential.

Avoid common mistakes on your manuscript.

1 Introduction

As artificial intelligence (AI) permeates our lives more and more, mechanisms such as deep neural networks [ 1 ] are put to use (or their deployment planned) in more and more places in various distributed systems. In particular, a wireless sensor network (WSN) can improve its coverage and connectivity and reduce energy and bandwidth usage by deploying AI onto edge nodes [ 2 ].

The data pipeline with AI typically requires the creation and the use of a model , i.e. a layered structure of complex algorithms (also known as operators ) which interpret data and make decisions based on that data. This model must first be trained ( learning phase [ 1 ]), before it can be put in production (used for inference ).

Recent work from the tiny machine learning (TinyML) [ 3 , 4 ] community forays into optimizing models to fit tinier resource budgets (and to perform efficiently nevertheless) on low-power microcontrollers in the Internet of Things (IoT). As a consequence, both learning and inference placement possibilities are extended to encompass ultra-low-power terminals.

However, generic and convenient open-source tools lack designers tackling a combination of AI and IoT (AIoT), who are required to do the following:

Evaluate the performance of their models when placed somewhere along the terminal-edge-cloud continuum, especially when including potential placement on different microcontroller-based devices

Fine-tune their models, and identify performance bottlenecks at model layer granularity, on different microcontrollers

Select an adequate microcontroller to execute their model, for a targeted task running on a low-power device to-be-designed

Continuously and securely update pre-provisioned models on fleets of heterogeneous devices under test, remotely, over the network

Monitor the computational performance of deployed models remotely, over the network, e.g. on new data collected in the field

This paper thus introduces the RIOT toolkit for machine learning (RIOT-ML), an open-source AIoT toolkit that tightly combines a generic model compiler and a popular low-power IoT operating system. RIOT-ML implements a workflow to automatically compress, flash and evaluate arbitrary models Footnote 1 on arbitrary commercial off-the-shelf (COTS) low-power boards Footnote 2 based on different popular microcontroller (MCU) architectures. By leveraging a widely applicable low-power network stack combined with secure IoT software mechanisms, RIOT-ML also provides the ability to control, monitor and update ML model on fleets of such devices remotely, over heterogeneous network topologies.

Paper contributions Our contributions are as follows:

We designed a universal toolkit for on-board evaluation and remote monitoring of TinyML models (RIOT-ML) on low-power devices. RIOT-ML integrates and extends the toolkit U-TOE [ 5 ], providing feasibility checks for the use of arbitrary models in a wide selection of IoT hardware platforms. It allows researchers and developers to locate the performance bottleneck of a given model on a target device. The evaluation results enable co-design with other components at system level and help optimize ML models and configurations for specific use cases, allowing to achieve the best possible performance on target devices.

We design, implement and evaluate mechanisms for secure over-the-air (OTA) model updates, as well as for continuous deployment and management of arbitrary models on resource-constrained devices over arbitrary network configurations, which may also include low-power network links.

We released the code Footnote 3 of RIOT-ML under an open-source licence. This implementation enables compilation, flashing, evaluation and secure OTA updates of various neural networks (computational graph-based models) from mainstream ML frameworks onto various low-power boards based on popular instruction set architectures (ARM Cortex-M, ESP32, RISC-V).

We provide benchmarks and a comparative experimental evaluation using RIOT-ML, reproducible both on an open-access IoT testbed and on personal workstations, which provide insights on inference performance with different models on different low-power hardware and demonstrate how RIOT-ML can be re-used by TinyML experimental researchers and developers to fine-tune IoT configurations.

2 Related work

Recent work has surveyed [ 6 ] the scope of ML frameworks, tasks and metrics, including a comprehensive review of the TinyML stack and deployment pipeline. A number of challenges need to be met in order to fit the tiny resource budgets typical of microcontrollers (kiloBytes of memory, power consumption in milliWatt, central processing unit (CPU) frequency in megahertz) while maintaining performance at an acceptable level and retaining portability to extremely polymorphic hardware in this category.

Embedded IoT software platforms

Various open-source IoT operating systems are used to provide hardware abstraction, resource-sharing primitives and convenient peripheral access (e.g. sensor/actuator, network subsystem) on heterogeneous low-power microcontrollers. Prior work such as [ 7 ] surveys such operating systems, among which prominent examples include RIOT [ 8 ] and Zephyr. Footnote 4 So far, however, such software platforms offer no support for ML framework—or if they do, this support is very limited. Moreover, the most advanced supports so far are typically hardware—or vendor-specific, e.g. with libraries provided by STM32CubeMx or ARM CMSIS-NN.

Low-power IoT testbeds

Various testbeds offer remote access to fleets of reprogrammable microcontroller-based devices. Prior work such as [ 9 , 10 ] survey such testbeds, among which prominent examples include the open-access facility IoT lab [ 11 ], which offers remote bare-metal access (serial over Transmission Control Protocol (TCP)) to a fleet composed of hundreds of popular low-power boards of various kinds.

Benchmarking suites for TinyML

Benchmarking ML on low-power hardware entails a number of challenges [ 12 ]. Prior work such as MLPerf Tiny [ 13 ] provides a standard benchmark suite (a fixed set of representative ML tasks) for evaluating the performance of given hardware and an online platform for manufacturers to publish their comparative benchmark results. In contrast, RIOT-ML offers a more powerful and more customizable toolkit for performing feasibility checks of user-defined machine learning models on low-power devices, with a greater degree of flexibility and customization.

TinyML benchmarks

Prior work such as [ 14 ] focuses on performance comparison of different machine learning frameworks on two COTS low-power boards (Arduino Nano BLE 33 and STM32 NUCLEO-F401RE). In particular, it benchmarked two TinyML frameworks, Tensorflow Lite for Microcontrollers (TFLM) and X-CUBE-AI over gesture recognition and wake word spotting. Other work such as [ 15 ] tested TFLM models on several microcontroller-based boards. While such papers provide a performance comparison of specific frameworks on specific boards for specific tasks, RIOT-ML offers greater flexibility and generality, allowing developers to evaluate a wider range of (user-specified) models on a larger variety of low-power devices, and to dive into the execution details of ML models.

TinyML model transpiler and compilers

Compilers such as Tensor Virtual Machine (TVM) [ 16 ] can be used to automate the transpilation and compilation of models provided by major ML frameworks (TFML, Pytorch, etc.) so as to expose low-level routines and optimize them for execution on specific processing unit characteristics (CPU, graphics processing unit (GPU) etc.). An extension of TVM called uTVM was recently introduced, adding smaller hardware targets including a variety of MCUs (microcontroller units).

TinyML model profilers

ML-EXray [ 17 ] enables TinyML developers to gain visibility into the layer-level details of ML execution and diagnose cloud-to-edge deployment issues. Developers can analyze and debug edge deployment pipelines with high usability, using less than 15 lines of code for a full examination. However, the reliance on Tensorflow Lite restricts the capability to accommodate models from further ML frameworks and deploy them on low-power devices. Major ML frameworks (TFML, Pytorch, MXNet etc.) provide internal profiler [ 18 ]. Such tools allow developers to measure the performance of their models. They can be used to collect metrics such as inference time and memory usage, which can then be analyzed to optimize the model’s performance. Though it can provide us execution details at layer level, it still lacks the support for on-device deployment and evaluation on various IoT devices, while RIOT-ML is a more general-purpose toolkit that provides a comprehensive solution on a wide range of low-power devices.

Tiny machine learning operations (TinyMLOps)

MLOps is a paradigm adapted from software engineering (Development Operations (DevOps)) and aims to deliver and maintain ML models efficiently and reliably in the field. It requires continuous end-to-end deployment of model and monitoring of its performance. TinyMLOps extends MLOps to support continuous integration and continuous deployment (CI/CD) workflows of ML model on resource-constrained devices. Several frameworks [ 19 , 20 , 21 ] provide MLOps components for conventional servers or large-scale clusters, but typically do not encompass TinyML devices due to the unique challenges these incur, in particular, extreme polymorphism in hardware architectures and network stacks. As a result, TinyMLOps (and to a large extent MLOps itself) is a paradigm that is still in its infancy.

Software updates for low-power IoT

Software updates in the context of IoT are crucial for resolving security vulnerabilities, extending functionalities and improving the performance of low-power devices during their lifetime. Research studies such as [ 22 , 23 , 24 ] focused on secure update mechanisms, including covering techniques for authentication, integrity checks and encryption, aiming to mitigate attacks coming from potential adversaries with various levels of computing power, up to quantum computing power levels. Efficiency and reliability aspects are explored, along with strategies for network load minimization and power management, aiming to optimize the OTA update process in IoT environments with limited network and battery resources [ 25 ]. Standardization efforts at the Internet Engineering Task Force (IETF), such as IPv6 over Low-Power Wireless Personal Area Networks (6LoWPAN) and Constrained Application Protocol (CoAP) [ 26 , 27 ], have shrunk on-board memory footprint and network transfer costs of Internet Protocol version 6 (IPv6) and Hypertext Transfer Protocol (HTTP)-like interaction over the network, while more recent standards such as Software Updates for Internet of Things (SUIT) [ 28 ] aim further at specifying generic architectures, data models and metadata for low-power secure IoT software updates over transfer protocols such as CoAP.

The above is summarized in Table 1 .

3 Background on TinyML performance analysis

On the one hand, as the most immediate limiting resource budget on microcontrollers concerns memory limitations, typically in the order of kiloBytes, TinyML performance evaluation typically focuses primarily on metrics measuring memory consumption—while keeping an eye on execution speed—as described below. On the other hand, TinyML performance analysis can be tackled at different granularity levels: at the global model level or at the operator level, for finer granularity, as described in the following.

3.1 Performance metrics

The considered metrics offer insights into the feasibility, efficiency and resource utilization of offloading model inference burden to low-power devices. By analyzing these metrics, users can make initial decisions regarding model selection, optimization techniques and hardware configurations to maximize performance and minimize the resource footprint on low-power devices.

Memory (RAM) consumption

This metric measures the amount of dynamic memory space (primary random access memory (RAM)) consumed by the model during inference. It reflects the memory footprint of the model activation and is important for low-power devices that have limited memory resources. Efficient memory utilization allows for the deployment of larger and more complex models on such devices.

Storage (flash memory) consumption

This metric quantifies the amount of storage space, typically in terms of flash memory region, required to store the compute instruction and associated parameters. It reflects the model’s storage footprint on the low-power device. Minimizing storage consumption allows for accommodating multiple models on the device or orchestrating with other essential applications.

Computational latency

This metric measures the time consumption of performing inference for each input sample, either at the model level or at the level of individual operators within the model. It reflects the inference speed of the model on the low-power device and plays a crucial role in real-time or latency-sensitive applications. Core clock frequency, cache strategies and communication latency between memory and working core have a great impact on this indicator.

This metric considers the cost of the System-on-a-Chip (SoC) used in the low-power device. The price of the SoC affects the overall affordability and feasibility of deploying model in large-scale distributed systems. Lower-cost SoCs can make the deployment more accessible and cost-effective.

3.2 Measurement granularity

As for performance analysis of machine learning in other domains, TinyML performance can be measured at different granularity levels:

Per-model evaluation

At this coarse level, one measures the performance of the model as a whole, i.e. the resource footprint incurred by the execution of the model including all its layers and operators. For example, this allows for evaluating the resource consumption for inference with the production-ready code, on a particular industrial hardware setup.

Per-operator evaluation

At this level, one measures separately the performance of one or more operators (i.e. one or more components of the model). This per-operator measurement can help identify specific operators that contribute to performance or inefficiencies, in optimizing the model’s efficiency and spotting potential bottlenecks.

4 Background on TinyML model update

A model deployed and running on a device can be updated at different levels. Depending on this level, flexibility and network traffic costs can vary.

Firmware updates

As surprising as it may be, the dominant approach in the domain of low-power IoT software updates on microcontroller-based devices is to perform a firmware update, i.e. transferring, verifying and (re)installing the totality of the software running on the device (except a minimal bootloader, in some cases). A simplistic approach to update a model remotely over the network is thus to roll out and deploy a firmware update containing the new model. This level requires transmitting the largest amount of data to the device can bloat network load and buffer memory requirements.

Full model updates

A more refined approach to updating software remotely over the network is to roll out and deploy fractional software updates instead of firmware updates. With this approach, a user updates only model binary code, i.e. its operators and related parameters (weights, bias etc.). Network transfer is decreased, while flexibility remains high, as a user can renew the operators with optimized execution flows, or even replace a model with a totally different architecture with better performance in accuracy or resource utilization.

Partial model updates

An even more fine-grained approach is to update only a subset of specified parameters of an already deployed model. For instance, a user can specify a single layer (or even: a single parameter) to be updated, which can minimize network traffic and related power consumption. Such fine-grained level may suit low-power IoT nodes with the same model exchanging parameters (federated/distributed learning) or receiving newer parameters from a central server. However, note that flexibility is decreased, as such an update cannot significantly alter a model’s architecture.

5 Threat model and security of TinyML model updates

Users trust machine learning models to provide accurate and unbiased predictions. To maintain model trustworthiness, at the very least, integrity checks and authorization during the model update process are necessary for ensuring the reliability of TinyML systems and to mitigate attacks such as model poisoning, where attackers attempt to manipulate training data or inject malicious bits into the model parameters.

In this paper, we do not consider sophisticated defenses against model poisoning. Rather, we consider five fundamental attack vectors during model update process:

Tampered model update

An adversary holds the model repository and tries to upload flawed models, which will be fetched by IoT devices for model update.

Tampered parameter update

This is the same as the above, but at the granularity of (malicious) parameter tilting.

Unauthorized model update

An unauthorized adversary attempts to request the IoT device to fetch and deploy modified models.

Unauthorized parameter update

This is the same as the above, but at the granularity of (unauthorized) parameter changes.

Break of confidentiality

The updates to models or parameters should be encrypted such that only the authorized maintainer and the device have access to the decrypted model updates.

Note that this threat model has limits: it does not cover the case where the root of trust, i.e. the authorized model maintainer, is itself compromised. For instance, the authorized maintainer himself could go rogue, or be otherwise tricked into lacing the model updates with malware, or into malicious tilting of model parameters. Nevertheless, any threat model must cover the above vectors, and in this sense, the security mechanisms we design are fundamental. They should allow extensions towards mitigating more sophisticated attacks.

6 RIOT-ML framework

RIOT-ML integrates uTVM and RIOT to perform model compilation, optimization and flashing, secure model update and model management and utilizes U-TOE for evaluation of arbitrary models produced by mainstream ML frameworks onto various low-power boards.

As depicted in Fig.  1 , users can use a host personal computer (PC) (combining Linux and the RIOT-ML toolchain) to deploy, evaluate and update their models on local IoT devices (e.g. connected via Universal Serial Bus (USB) to their PC), or on remote IoT devices via a testbed offering bare-metal access to various types of MCUs over the network.

Hardware setup of RIOT-ML. Users can connect local boards with host PC via serial or use remote board service on IoT testbed

6.1 Architectural design

As depicted in Fig.  2 , the toolkit is composed of the following key components:

Model compiler : RIOT-ML leverages the uTVM compiler to convert arbitrary neural network models into efficient C code. This compiler takes as input the output of typical ML frameworks (such as PyTorch, TensorFlow...), then enhances the efficiency of the models for microcontrollers and enables them to be run on low-power devices.

OS environment and hardware support : RIOT is a general-purpose OS for low-power IoT devices, which was chosen to provide a lightweight runtime environment for model execution and evaluation on microcontrollers. This base provides extensibility and wide-spectrum support for heterogeneous low-power boards.

Model evaluation module : This component integrates the U-TOE toolkit [ 5 ] to perform on-board model evaluation. Section  9 describes its architecture and measurement procedures in detail.

Secure model update module : This module integrates an implementation of Request for Comments (RFC) 9019 (SUIT [ 22 ]) to provide a secure, low-power software updates on IoT devices. It utilizes SUIT manifests, which contain model version information, payloads’ hashes and digital signatures from maintainers, to ensure the integrity and authenticity during update process. We adapted and integrated it into RIOT-ML to support secure OTA model update as described in Section  6.2 .

Network management endpoints : RIOT-ML exposes CoAP endpoints on IoT devices running the framework, serving as interfaces for model management. Users can query model metadata and status (name, version, evaluation results etc.), control model behaviours or trigger model update, as described in Section  6.2 .

Fleet management module : RIOT-ML implements fleet management on the host for remote updates of models, automating a build pipeline to support CI/CD of machine learning models on device. This model builds update payloads, generates and signs SUIT manifests, pushes them to a repository (CoAP registry) and notifies the managed IoT devices about model updates availability. It also provides a user-friendly interface.

Additionally, RIOT-ML provides a connector for a cloud-based IoT testbed which enables seamless interaction with remote boards using serial over TCP.

Software architecture and components of the RIOT-ML framework. In grey: U-TOE toolkit components

Compilation and deployment workflow of RIOT-ML. uTVM optimizes and translates model from mainstream ML framework into model library, which is co-compiled and flashed with RIOT and U-TOE components onto target boards

We depict in Fig.  3 the high-level view of the typical workflow with RIOT-ML. In a preliminary step, RIOT-ML first gathers the specification of target device to decide the compilation options for uTVM and RIOT. Then, uTVM generates a non-optimized model “library”. Some static optimization strategies are then applied in this stage, according to the target device type, in particular, to determine an appropriate scheduling. Footnote 5 This uTVM-generated library is then jointly compiled with RIOT, an RPC server and a measurement worker into an executable firmware, which is then automatically flashed on the device (via USB, or remotely via the network).

6.2 Model over-the-air update and management

We provide interfaces and modules both on the IoT devices and on the host (fleet manager) to enable secure OTA update and management.

Architecture

On the one hand, software components embarked on IoT devices which can answer to management commands and act upon update notifications were implemented on top of RIOT’s low-power IPv6 network stack and a SUIT implementation. Update and management interfaces were implemented as separate RIOT modules. The host on the other hand, as fleet manager, is responsible for building and coordinating updates, sending commands and notifications and maintaining an update repository (a server with a CoAP registry) which stores update payloads.

Model management

To remotely control, maintain and monitor ML model operation and performance efficiently and continuously, RIOT-ML exposes several CoAP endpoints on managed IoT devices, as shown in Table  2 . These endpoints can be accessed by the authorized maintainer via the fleet manager module.

Model update procedure

The procedure consists of two steps. In the first step, depicted in Fig.  4 , an authorized maintainer identified with the public/private key pair \((P_k, S_k)\) produces and pushes the update binary and the associated metadata secured as per SUIT (the SUIT manifest). In a second step, depicted in Fig.  5 , the maintainer (using the fleet manager module) notifies the managed IoT devices that a new update should be fetched, verified and installed, which then ensues.

Note that, as per the SUIT specification, different schemes for digital signatures, hashing and encryption can be used (performance comparisons are available in prior work such as [ 24 ]). In the below, we assume the use of Secure Hash Algorithm 256-bit (SHA256) [ 29 ] for hashing, and Ed25519 algorithm [ 30 ] for digital signatures.

The following describes the update procedure in more detail:

Preliminary : Pre-provision the device with the authorized maintainer’s public key \(P_k\) .

The fleet manager encapsulates the model binary (or the selected model parameters to be updated) into a SUIT payload format and computes the corresponding SHA256 digest as payload checksum.

The fleet manager generates and signs the SUIT manifest using an Ed25519 signature scheme, which contains the uniform resource identifier (URI) and checksum of the payload.

The fleet manager pushes the update binary and the SUIT manifest to the CoAP server (matching the URI).

Through a dedicated exposed CoAP endpoint, the device is notified with the URI of the new SUIT manifest.

The device fetches and verifies (using \(P_k\) ) the SUIT manifest. If the verification fails, the update aborts and reports an error code.

The device fetches the update binary (payload) designated in the SUIT manifest and performs integrity check. If the verification fails, the update aborts and reports an error code.

The device installs the new model in RAM or in Flash memory, depending on the part of the model that is updated.

Generation and pushing of SUIT manifest and payloads for secure model update with RIOT-ML. The key pair ( \(P_k\) , \(S_k\) ) authenticates the authorized maintainer. The CoAP server has no knowledge of secret key \(S_k\) and only serves as artefacts repository

6.3 Security guarantees with RIOT-ML over-the-air updates

To counter the potential attacks using model update as a vector, RIOT-ML uses SUIT to provide the following guarantees:

Model update integrity

RIOT-ML inspects the checksums of the update payloads and the metadata specified by SUIT, signed by the model maintainer. This ensures the integrity and prevents tampered update payload.

Model update authenticity

By using the combination of a digital signature mechanism and a cryptographically secure digest specified by SUIT, the RIOT-ML framework guarantees that only the users with the valid private keys have the authority to update the model.

Notification, fetching, verification and installation of model updates with RIOT-ML, securely over the network

Model update confidentiality

The IETF also specifies (op-tional) encryption mechanisms for SUIT payload binaries (see [ 31 ]). For this, a symmetric content encryption key is used, this key being either pre-shared or established on-the-fly for instance via an ephemeral-static Diffie-Hellman exchange (EDHOC) [ 32 ]. If used, this mechanism provides confidentiality for the updates of models.

7 U-TOE toolkit

U-TOE is a toolkit we designed and implemented, integrating RIOT and uTVM to perform model compilation and evaluation, on-board a microcontroller-based device.

7.1 Architectural design

The toolkit is composed of the following key components:

RPC module : To evaluate the resource consumption at operator level, U-TOE utilizes the Remote Procedure Call (RPC) mechanism of uTVM. The RPC mechanism enables to upload and launch functions onto IoT boards over serial. This is useful for remote testing and profiling, enabling U-TOE to wrap the model operators for measuring the computational latency and memory usage. It is composed of a client on the host and a server on the target device, receiving commands and executable instructions from the host.

Evaluation module : It contains two units: measurement worker and analyzer. As shown in Fig.  2 , the measurement worker is deployed on the MCU for acquiring performance metrics at model or operator level. Besides carrying out the measurement of resource footprint, it is in charge of the randomization of model input and is responsible for reporting metrics data to host device. The analyzer runs on the host, statisticizes the uploaded metrics from device and provides a human-readable frontend for users.

Evaluation workflow with U-TOE

: After deploying the program on the target device, a bidirectional channel is set up between host and device, as depicted in Fig.  2 . The measurement worker starts collecting performance metrics at user-specified level and uploads metrics data to the analyzer. Eventually, users can obtain the overall statistics of model metrics, or catch the performance bottleneck with execution details of each operator. All the raw metrics data are saved in a log file for further, user-customized analysis.

7.2 Measurement procedure

We designed two measurement procedures to support evaluation at different granularity. The procedures run across multiple components of the toolkit and most workloads are primarily on the target board. The following steps describe the measurement routine after the compilation of an executable program. The steps marked with bold number are executed on the target board.

This mode focuses on the model performance in an actual production environment. Here is the corresponding measurement routine:

Calculate model memory and storage consumption based on Executable and Linkable Format (ELF) file. We disable dynamic memory allocation to enable static analysis of memory footprint.

Deploy executable program to local or remote IoT board.

Repeat model inference based on the user-specific number of trials with randomized input on uniform distribution.

Record computational latency of each trial.

Upload records to host device for further analysis and archive.

At the end of evaluation, results with statistics (e.g. 95% confidence interval, median, maximum and minimum) are presented on the host device, including computational latency and consumption of memory and storage.

In contrast to per-model evaluation, this mode focuses on the efficiency and resource footprint of each operator, enabling to discover the performance bottleneck inside models. Thanks to the time evaluator inside the uTVM’s RPC mechanism, we can measure computational latency at operator level. The high abstraction of timer and serial in RIOT allows us to unify the implementation of time measurement and RPC communication on arbitrary IoT boards. Here is the corresponding measurement routine:

Analyze memory footprint at operator level utilizing internal API of TVM.

Start RPC Server on IoT board.

Launch RPC client to benchmark and record the execution performance of each operator.

It is noted that the operator structure constructed by uTVM is usually inconsistent with the hand-crafted version in the ML framework. That is because uTVM as model compiler applies model optimization (i.e. operator fusion) and inserts execution details (i.e. quantization arithmetic) during conversion and compilation, which potentially merges multiple operators into a single one or inserts additional operators. Nevertheless, we annotate the operators from uTVM with model parameters (weights, biases etc.), so that users can associate a specific operator to the corresponding layer.

8 Experiments with RIOT-ML

We next demonstrate the capabilities of RIOT-ML. We report on two categories of experiments that were carried out. In this section, we first describe the experimental setup for each category of experiments, then Section  9 will present and analyze the results we gathered.

8.1 Model performance evaluation

We conducted experiments to validate the functionality and compatibility of RIOT-ML on model side (universal support for model structure and ML frameworks) and on device side (wide-spectrum support for IoT devices). Hence, we dived into two orthogonal directions: For device support, we evaluated a quantized LeNet-5 on various IoT boards; for model compatibility, we evaluated multiple models on a local STM32F746G discovery board.

Model selection

We selected pre-trained, quantized models from open-source repositories, Footnote 6 which target on typical TinyML tasks (image classification, keyword spotting, visual word wake, noise suppression and abnormal detection). The weights and activations of the model were quantized to 8-bit integer, yet the inputs and outputs remain in IEEE 754 single-precision floating point [ 33 ] format.

Model optimization

We only used built-in, rule-based optimization in uTVM. Thus, all heuristic optimization strategies like model scheduling were disabled.

MCU Configuration

We disabled data and instruction cache to observe the “memory wall” effect in ML model. The core clock frequency was preset by CPU initialization code in RIOT and is presented with experiment results in Section  9 .

Hybrid deployment

The experiments were conducted both on local and remote IoT boards provided by FIT IoT-LAB.

It is noted that for each evaluation, we preset the number of trials to ten in order to address random error.

8.2 Secure model OTA updates

We conducted experiments updating models running on a nRF52840dk Development Kit board in the FIT IoT-LAB testbed and measured the resource consumption of significant components and the network transfer overhead. Our measurements used the LeNet-5 model as a use case. We consider different granularity for the update: either updating the firmware or, instead, a partial update concerning only the weights of the final layer (serving as a classifier) of the quantized LeNet-5 model.

Network setup

The nRF52840 board is connected to the network through a low-power IEEE 802.15.4 radio access link (using the Generic Network Stack (GNRC) in RIOT) on which 6LoWPAN, IPv6 traffic can flow. Via an intermediate IPv6 border router, CoAP messages can transit to/from the device and reach the fleet manager hosted on the user’s PC connected to the Internet. By using this approach (6LoWPAN and IPv6), we ensure that the approach can run end-to-end over the Internet over an arbitrary set of links which can include not only typical Ethernet and WiFi links, but also some low-power radio links (e.g. Bluetooth Low Energy (BLE), Long Range (LoRa), IEEE 802.15.4).

Crypto configuration

We produced public and secret keys using Ed25519 algorithm before commissioning the model update modules. CBOR Object Signing and Encryption (COSE) [ 34 ] was used as specified by SUIT to sign the manifest and CoAP payloads. The crypto libraries we used in practice are libcose Footnote 7 and c25519, Footnote 8 which target on low-memory systems and implement COSE and Ed25519, respectively.

9 Analyzing RIOT-ML measurements

In this section, we present the results of the experiments we describe in Section  8 .

9.1 Model performance evaluation

Table 3 presents the resource consumption of the LeNet-5 model on various IoT boards, generated by per-model evaluation. MCUs are grouped by family and arranged in ascending order of clock frequency within each group. ARM Cortex-M series MCUs showed no significant difference in memory and storage usage, and the computational latency declined as the core frequency increased. One outlier is the RP2040-based rpi-pico board, but the extra 16KB RAM is in fact reserved for the debugger. Benefits from the full support of digital signal processing (DSP) and Thumb-2 instruction set, Cortex-M3, -M4 MCUs perform better than Cortex-M0+ with the same core clock frequency. Another outlier was discovered on SiFive RISC-V MCU. With the highest core clock frequency, this won the least favorable ranking on computational latency and memory usage. This MCU uses an external, Serial Peripheral Interface (SPI) NOR flash for data and program storage, causing a huge performance regression while we disabled the cache.

Table 4 presents the results of various ML models on representative TinyML tasks on individual IoT boards, showcasing the universal support for various ML frameworks and model structures. Except for LeNet-5 trained on a local host device with Pytorch, all the others came from open-source model zoos. Memory and storage columns refer to their resource consumption.

Unsurprisingly, we observe how storage consumption decreases proportionally to decreasing the number of model parameters. One apparent outlier is DS-CNN Small , which consumes more storage space than LeNet-5 , although it has nearly twice the number of model parameters. Further analysis revealed however that DS-CNN Small contains almost three times more scaling factors (in floating point) compared to LeNet-5 . Those scaling factors are not counted as model parameters but occupy a large amount of storage.

As expected, the most complex model, MobileNetV1 , consumed the most memory (RAM) and computation resources. However, as shown in Table 4 , the execution overhead, memory consumption and computational latency cannot be reliably predicted solely from the amount of model parameters in general. The model structure and its associated computational pattern, as well as the effects of compression, also impact intensely on the execution overhead (hence the usefulness of RIOT-ML as an experimental toolkit for benchmarking!).

We here used a tiny model with only three layers from TFlite as an example to avoid unnecessary complexity in demonstration, with output results presented in Table 5 . The computational bottlenecks are located in operator add_nn_relu and add_nn_relu_1 , and with the highest memory and storage consumption as well. We can trace down the corresponding layers of the original model with the hints of associated parameters, which are the weights, bias or other trainable parameters of the model, making it possible to apply optimization strategies on well-targeted layers.

9.2 Secure model OTA updates

We show in Table  6 a breakdown of resource consumption. Since full and partial updates share the same software stack, their consumption remains identical. We observe that due to its large number of constant model parameters (weights and bias), the ML model is the component which uses the most storage (flash memory). Concerning (RAM) memory usage, the network stack and SUIT module use relatively more memory because of the large buffer requirements for network packets and SUIT payloads handling. We also observe that the cryptographic implementations employed by SUIT (namely libcose and c25519) claim minimal memory and storage consumption.

9.3 Preliminary assessment of network transfer costs

We additionally measured in Table  6 the size of SUIT manifest and update binary payloads which must be transmitted over the network to carry out the model update process. These measurements can help gauge the incurred load on the network. Concerning the payloads, the serialized firmware’s size and the weights of the final layer are 123.44 KB and 0.84 KB, respectively. Comparatively, the SUIT manifest’s size is 0.471 KB, which represents a rough ratio of 260:2:1. This means in particular that the overhead on network load incurred by using SUIT security mechanisms in RIOT-ML (for integrity, authentication and authorization) remains well under 1% when the approach of firmware updates is used to update the models.

Discussion on model update granularity

The simplest app-roach to model update over-the-air on IoT devices is firmware updates, as the embedded system is relieved of the complexities stemming from dynamic loading on heterogeneous hardware. However, this convenience comes with a high price in terms of network transfer costs. The measurements in Table  6 hint that updating only the model (49 KB) instead of the full firmware (126 KB) would save more than 60% off the network transfer costs. Going further, if in some cases model updates could get away with transfer learning, i.e. updating only parts of the model network transfer costs become comparatively negligible. In our measurements, for instance, we updated only the last layer, which eliminated more than 99% of the network transfer costs. However, with this approach, inference accuracy improvements with the update may be more limited. Thus, users are obliged to balance between network overheads, technical complexity and model accuracy when designing their model update scheme.

10 Reproducible and custom RIOT-ML experiments

We released the full source code of the RIOT-ML toolkit on Github at https://github.com/TinyPART/RIOT-ML under an open-source LGPL v3 licence. For further details on how to start with RIOT-ML hands-on, the reader is referred to the comprehensive Readme.md in the repository.

On the one hand, researchers and practitioners who possess IoT hardware that is supported by the open-source operating system RIOT (currently 250+ types of boards, using 60+ types of CPUs) Footnote 9 can use RIOT-ML out-of-the-box, directly on their boards.

On the other hand, combined with the use of the free open-access testbed IoT-Lab, Footnote 10 even researchers and practitioners who do not have such hardware on-premises can conduct large-scale experimental evaluation campaigns using RIOT-ML.

Perspectives

As RIOT board and CPU support expands and improves over time, and as uTVM also expands support to other architectures in parallel (both open-source communities are very active), RIOT-ML can in a very short time expand its support for new use cases, automatically adding the support of uTVM for new boards, and the support of RIOT for new models. As such, RIOT-ML may organically grow and become a useful link between the two communities.

Moreover, while the work on RIOT-ML in this paper has been focused on inference only on single-core microcontrollers, there is strong potential to extend the toolkit provided by RIOT-ML to support on-device learning scenarios and optimize the exploitation of multi-core microcontrollers.

11 Conclusion

This paper introduces RIOT-ML, a novel toolkit we designed to streamline the endeavours of AIoT practitioners. RIOT-ML facilitates TinyML model continuous integration, secure continuous deployment and performance evaluation, remotely over the network, on fleets of heterogeneous IoT devices under test. More precisely, with RIOT-ML, users can take the model zoo output by various traditional machine learning frameworks (such as PyTorch, TensorFlow) and automate their adapting, their deployment and the assessment of their performance when executed on a wide selection of IoT boards and development kits based on various types of low-power microcontrollers (ARM Cortex-M, ESP32, RISC-V). Enabling and facilitating such large test matrices is indeed crucial for advancing the field of AIoT. To this end, we also published a highly reusable, well-documented, and customizable open-source implementation of RIOT-ML, which leverages the vibrant open-source communities associated with RIOT and uTVM. Last but not least, we showcase the practical application of RIOT-ML by providing preliminary experimental evaluation results on an open-access testbed.

Data availability

No datasets were generated or analyzed during the current study.

Output of TensorFlow, PyTorch...

Such as BBC:microbit, nrf52840dk, Arduino Zero, HiFive...

see https://github.com/TinyPART/RIOT-ML

See https://www.zephyrproject.org/

A schedule specifies low-level optimization for loop execution, enhancing cache hit and memory access. The optimal schedule is co-determined by model and device specification and identified by heuristic search algorithms based on measurements on device [ 16 ].

see https://github.com/ARM-software/ML-zoo and https://github.com/mlcommons/tiny

See https://github.com/bergzand/libcose

See https://www.dlbeer.co.nz/oss/c25519.html

see https://github.com/RIOT-OS/RIOT/tree/master/boards

See https://www.iot-lab.info/

Sze V, Chen Y-H, Yang T-J, Emer JS (2017) Efficient processing of deep neural networks: a tutorial and survey. Proc IEEE 105(12):2295–2329

Article   Google Scholar  

Sharma H, Haque A, Blaabjerg F (2021) Machine learning in wireless sensor networks for smart cities: a survey. Electronics 10(9):1012

Sanchez-Iborra R, Skarmeta AF (2020) TinyML-enabled frugal smart objects: challenges and opportunities. IEEE Circuits Syst Mag 20(3):4–18

Ray PP (2022) A review on TinyML: state-of-the-art and prospects. J King Saud Univ Comput Inf Sci 34(4):1595–1623

Google Scholar  

Huang Z, Zandberg K, Schleiser K, Baccelli E (2023) U-TOE: universal TinyML on-board evaluation toolkit for low-power IoT. In 2023 12th ifip/ieee inter-national conference on performance evaluation and modeling in wired and wireless networks (pemwn), IEEE, pp 1–6

Saha SS, Sandha SS, Srivastava M (2022) Machine learning for microcontroller-class hardware-a review. IEEE Sensors J 22(22):21362–21390

Hahm O, Baccelli E, Petersen H, Tsiftes N (2015) Operating systems for low-end devices in the Internet of Things: a survey. IEEE Internet Things J 3(5):720–734

Baccelli E, Gündoğan C, Hahm O, Kietzmann P, Lenders MS, Petersen H, Wählisch M (2018) RIOT: an open source operating system for low-end embedded devices in the IoT. IEEE Internet Things J 5(6):4428–4440

Lima LE, Kimura BYL, Rosset V (2019) Experimental environments for the Internet of Things: a review. IEEE Sens J 19(9):3203–3211

Gluhak A, Krco S, Nati M, Pfisterer D, Mitton N, Razafindralambo T (2011) A survey on facilities for experimental Internet of Things research. IEEE Commun Mag 49(11):58–67

Adjih C, Baccelli E, Fleury E, Harter G, Mitton N, Noel T, Vandaele J et al (2015) FIT IoT-LAB: a large scale open experimental IoT testbed. In: 2015 IEEE 2nd world forum on internet of things (wf-iot), IEEE, pp 459–464

Banbury CR, Reddi VJ, Lam M, Fu W, Fazel A, Holleman J, Lokhmotov A et al (2020) Benchmarking TinyML systems: challenges and direction. Preprint at arXiv:2003.04821

Banbury C, Reddi VJ, Torelli P, Holleman J, Jeffries N, Kiraly C, Pau D et al (2021) MLPerf tiny benchmark. Preprint at arXiv:2106.07597

Osman A, Abid U, Gemma L, Perotto M, Brunelli D (2022) TinyML platforms benchmarking. In: Applications in electronics pervading industry, environment and society: Applepies 2021, Springer pp 139–148

Sudharsan B, Salerno S, Nguyen D-D, Yahya M, Wahid A, Yadav P, Ali MI (2021) TinyML benchmark: executing fully connected neural networks on commodity microcontrollers. In: 2021 IEEE 7th world forum on internet of things (wf-iot), IEEE, pp 883–884

Chen T, Moreau T, Jiang Z, Zheng L, Yan E, Shen H, Ceze L et al (2018) TVM: an automated end-to-end optimizing compiler for deep learning, 578–594

Qiu H, Vavelidou I, Li J, Pergament E, Warden P, Chinchali S, Katti S (2022) ML-EXray: visibility into ml deployment on the edge. Proc Mach Learn Syst 4:337–351

Yousefzadeh-Asl-Miandoab E, Robroek T, Tozun P (2023) Profiling and monitoring deep learning training tasks. In: Proceedings of the 3rd workshop on machine learning and systems, pp 18–25. https://doi.org/10.1145/3578356.3592589

Kreuzberger D, Kühl N, Hirschl S (2023) Machine learning operations (MLOps): overview, definition, and architecture. IEEE Access 11:31866–31879. https://doi.org/10.1109/ACCESS.2023.3262138

Lê MT, Arbel J (2023) TinyMLOps for real-time ultra-low power MCUs applied to frame-based event classification. In: Proceedings of the 3rd workshop on machine learning and systems, New York, NY, USA: Association for Computing Machinery, pp 148–153

Diaz-de Arcaya J, Torre-Bastida AI, Zárate G, Miñón R, Almeida A (2023) A joint study of the challenges, opportunities, and roadmap of MLOps and AIOps: a systematic survey. ACM Comput Surv 56(4):1–30. https://doi.org/10.1145/3625289

Zandberg K, Schleiser K, Acosta F, Tschofenig H, Baccelli E (2019) Secure firmware updates for constrained IoT devices using open standards: a reality check. IEEE Access 7:71907–71920

Mtetwa NS, Tarwireyi P, Abu-Mahfouz AM, Adigun MO (2019) Secure firmware updates in the Internet of Things: a survey. In 2019 international multidisciplinary information technology and engineering conference (imitec), IEEE, pp 1–7

Banegas G, Zandberg K, Baccelli E, Herrmann A, Smith B (2022) Quantum-resistant software update security on low-power networked embedded devices. In International conference on applied cryptography and network security, Springer, pp 872–891

Bauwens J, Ruckebusch P, Giannoulis S, Moerman I, De Poorter E (2020) Over-the-air software updates in the Internet of Things: an overview of key principles. IEEE Commun Mag 58(2):35–41

Bormann C, Castellani AP, Shelby Z (2012) CoAP: an application protocol for billions of tiny Internet nodes. IEEE Internet Comput 16(2):62–67

Shelby Z, Hartke K, Bormann C (2014) RFC 7252: The constrained application protocol (CoAP). IETF Request for Comments

Moran B, Tschofenig H, Brown D, Meriac M (2021) A firmware update architecture for Internet of Things. RFC 9019. https://doi.org/10.17487/RFC9019

Hansen T 3rd, D. E. E. (2006) US secure hash algorithms (SHA and HMACSHA). RFC 4634. https://doi.org/10.17487/RFC4634

Josefsson S, Liusvaara I (2017) Edwards-curve digital signature algorithm (EdDSA). RFC 8032. https://doi.org/10.17487/RFC8032

Tschofenig H, Housley R, Moran B, Brown D, Takayama K (2023) Encrypted payloads in SUIT Manifests (Internet-Draft No. draft-ietf-suit-firmwareencryption-18). Work in Progress. Internet Engineering Task Force. Retrieved from https://datatracker.ietf.org/doc/draft-ietfsuit-firmware-encryption/18/

Selander G, Mattsson JP, Palombini F (2024) Ephemeral Diffie-Hellman Over COSE (EDHOC). RFC 9528. https://doi.org/10.17487/RFC9528

Kahan W (1996) IEEE standard 754 for binary floating-point arithmetic. Lecture Notes Status IEEE 754(94720–1776):11

Schaad J (2017) CBOR object signing and encryption (COSE). RFC 8152. https://doi.org/10.17487/RFC8152

Download references

Acknowledgements

The authors would like to thank Cedric Adjih, Nadjib Achir and Felix Biessmann for useful discussions and suggestions.

Open Access funding enabled and organized by Projekt DEAL. The research leading to these results partly received funding from the MESRI-BMBF German/French cybersecurity program under grant agreement No. ANR-20-CYAL-0005 and 16KIS1395K.

Author information

Authors and affiliations.

Freie Universität Berlin, Berlin, Germany

Zhaolan Huang, Koen Zandberg, Kaspar Schleiser & Emmanuel Baccelli

Inria Saclay, Paris, France

Emmanuel Baccelli

You can also search for this author in PubMed   Google Scholar

Contributions

Z. H. and E. B. conceptualized and wrote the main manuscript text; Z. H., K. Z. and K.S. developed and tested the software. All authors reviewed the manuscript.

Corresponding author

Correspondence to Zhaolan Huang .

Ethics declarations

Conflict of interest.

The authors declare no competing interests.

The paper reflects only the authors’ views. MESRI and BMBF are not responsible for any use that may be made of the information it contains.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Huang, Z., Zandberg, K., Schleiser, K. et al. RIOT-ML: toolkit for over-the-air secure updates and performance evaluation of TinyML models. Ann. Telecommun. (2024). https://doi.org/10.1007/s12243-024-01041-5

Download citation

Received : 17 January 2024

Accepted : 12 May 2024

Published : 22 May 2024

DOI : https://doi.org/10.1007/s12243-024-01041-5

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Machine learning
• Microcontroller
• Software update
• Find a journal
• Publish with us
• Track your research

• Facts & Figures
• Accreditations
• Maps and Directions
• Faculty Distinctions, Awards and Honors
• Engineering Honors
• Computer Engineering
• Global Programs
• Student Organizations
• Peer Teachers
• Fast Track: Optimizing the transition from Undergraduate to Graduate Studies
• Admissions and Aid
• Entry to a Major Process
• Scholarships and Financial Aid
• Research Areas
• Undergraduate Research
• Seminars and Distinguished Lectures
• Industry Capstone Program
• Industrial Affiliates Program

Online Artificial Intelligence and Machine Learning Certificate

Gain a competitive edge with our graduate-level Artificial Intelligence and Machine Learning Certificate. This program equips both novices and seasoned professionals with the essential skills to harness the power of modern Artificial Intelligence and Machine Learning in their domain. Upon completion, participants will master statistical analysis and machine learning techniques, enabling them to dissect complex data sets. Armed with the ability to synthesize and evaluate AI models, graduates will confidently tackle real-world challenges, leveraging cutting-edge tools to derive actionable insights and drive innovation in their respective fields.

I'm ready to apply!   Request more information

Certificate Overview

The Artificial Intelligence and Machine Learning certificate is a 12-credit program that equips novices and seasoned professionals with the essential skills to harness the power of modern Artificial Intelligence and Machine Learning in their respective fields of operation.

Technical Qualifications

To be successful in this program, prospective students must demonstrate an understanding of core concepts in computer science or equivalent covered in the categories below:

• Program Design and Concepts : programming proficiency through problem-solving with a high-level programming language, emphasizing computational thinking, data types, object-oriented design, dynamic memory management, and error handling for robust program development.
• Data Structures : implementing essential abstract data types and algorithms covering stacks, queues, sorting, searching, graphs, and hashing; examining performance trade-offs, analyzing runtime and memory usage.
• Algorithms : computer algorithms for numeric and non-numeric problems; design paradigms; analysis of time and space requirements of algorithms; correctness of algorithms.
• Discrete Structures for Computing : foundations from discrete mathematics for algorithm analysis, focusing on correctness and performance; introducing models like finite state machines and Turing machines.
• Mathematical Foundations : Calculus, Probability, and Linear Algebra.

Students must take four out of five possible courses to complete this certificate. See course information below.

Information

To qualify for this certificate, you must complete 12 semester credit hours (SCH) of coursework from the following list of courses. All courses must be completed with a grade of C or above. Each course is linked to its course description within the catalog.

Courses (12 credits):

Select four of the following:*

• CSCE 625 - Artificial Intelligence
• CSCE 633 - Machine Learning
• CSCE 635 - AI Robotics
• CSCE 636 - Deep Learning
• CSCE 642 - Deep Reinforcement Learning

* Additional courses are available with the consultation of an academic advisor.

For more information, please see the course catalog .

Why choose Engineering Online

Advance your career with our Engineering Online program! Backed by the university's esteemed reputation and national recognition in engineering education, you'll engage directly with industry leaders and a rigorous curriculum. Beyond graduation, tap into the extensive Aggie Alumni Network, offering invaluable connections to propel your career forward.

Engineering Online Benefits

Certificate Highlights

Related academics.

Online Master of Computer Science

Online Master of Engineering in Computer Engineering

Frequently asked questions.

Discover answers to frequently asked questions tailored to assist you in making informed decisions regarding your education with Engineering Online.

Graduate Admissions

Use EngineeringCAS to apply for the distance education version of the certificate. Follow the provided instructions, as they may differ from certificate to certificate.

Graduate Tuition Calculator

To calculate cost, select the semester you’ll start, choose “Engineering” from the drop-down menu, and slide “Hours” to how many you’ll take each semester. Your total cost is Tuition and Required Fees + Engineering Program Fee (Remote).

Questions? Email [email protected] !