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

Abstract

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.

Graphical Abstract

Open in new tabDownload slide

Introduction

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 CO₂ 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].

1 Methods and challenges addressed

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.

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 AI techniques

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.

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].

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].

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.

• 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

else

food_source <- random_food_source ()

Initialization phase:

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

where $xmax$ j is the upper bound of the parameter j and $xmin$ 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 $vij$ representative of a location such as a food source in her $xij$ 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 $xk$ is a randomly picked solution, j is a parameter randomly selected and $βij$ is a random integer within the [–a, a] range. A greedy selection between $vi$ and $xi$ is applied after the production of a new solution $vi$⁠.

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 $xi$ based on pi are picked to assess the new solutions $vi$ and new solutions $vi$ for spectators are created. The hired bees between $vi$ and $xi$ 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 $Vi$⁠, $Y∗(t)$ 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.

C₁ and C₂ are the learning parameters that determine the influence of the personal best and global best solutions, respectively.

r_and1 and r_and2 are randomly generated numbers within the range of [0,1].

P_best_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_best_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.

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. X1ij,..., 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), u0j,..., u1j represent the random effects. W1j and W2j 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 Application of AI in the power sector

3.1 Operation of the power system

3.1.1 ELD

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.

3.1.3 UC

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.

3.2 Control of the power system

3.2.1 VC

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.

3.3 Planning of power system

3.3.1 DSP

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.

4 Conclusions

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.

5 Future scope

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.

Authors’ contributions

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

Conflict of interest statement

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.

Data Availability

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

References