A systematic literature review on machine learning applications at coal-fired thermal power plants for improved energy efficiency

ABSTRACT

Power generation comprises high environmental and ecological impacts. The global power industry is under pressure to develop more efficient ways to operate and reduce the impacts of inherent process variability. With the rapid development of technologies within the energy sector, large volumes of data are available due to in-time operational measurements. With increased computer processing capabilities, machine learning is applied to explore these in-time operational measurements for improved process understanding. This research paper investigates machine learning algorithms for energy efficiency improvement at coal-fired thermal power plants by conducting a systematic literature review. This research is essential since it provides guidelines for applying machine learning towards sustainable energy supply and improved decision-making. Subsequently, efficient processes result in the reduction of fuel usage, which results in lower emission levels for equivalent power generation capacity. Furthermore, this study contributes towards future research by providing valuable insights from academic and industry-related studies.

KEYWORDS:

• Coal-fired thermal power plant
• energy efficiency
• machine learning
• systematic literature review

1. Introduction

The global power industry is under pressure to develop more efficient ways to operate and to reduce the impact of variable operations on market supply and the environment (Sleiti, Kapat, and Vesely 2022). Minimal attention was given to energy management during the early 1970s (Petrecca 2014). However, energy security was a big concern after the oil crisis, leading to more energy-efficient technologies and operations (Kaya and Keyes IV 1980). Studies published over the last few decades focused predominantly on energy management, emphasising the need for more effective ways to generate and use energy resources globally (Bunse et al. 2011; Narisco and Martins 2020; Schulze et al. 2016). This need originates from the challenges that emerge by balancing the use of valuable resources, the human population and providing proper living conditions (Narisco and Martins 2020).

Energy efficiency is the attempt to decrease the recourses used while considering resource depletion, monetary savings, environmental impacts, and resource security (Lovins 2017; Patterson 1996). Energy efficiency is considered one of the primary contributors to a sustainable society and is considered a critical aspect in all industries and countries. In South Africa, 2003 data indicate that the mining sector consumes more than 175 PJ of energy annually. Thus, it is identified as the highest consumer of electricity (Oladiran and Meyer 2007). These energy-intensive operations could be more sustainable, provided the government's regulations on carbon emissions (Awuah-Offei 2016).

Numerous initiatives have been launched globally over the last few years to highlight the importance of energy efficiency (Narisco and Martins 2020), including, but not limited to, climate and energy initiatives by the European environment agency (Climate & Energy Package 2020), and the energy star programme (Energy Star Overview 2016). Given this high electricity demand, the Chinese government and various other countries, including India (Mishra 2004) and Zimbabwe (Murehwa et al. 2012), focus explicitly on thermal power plants for electricity generation. A thermal plant consists of various boilers that feed into common steam receivers (Murehwa et al. 2012). These feed sources are mainly coal, oil, nuclear or gas. Most power plants use coal to generate electricity (Bi et al. 2014; Murehwa et al. 2012).

A coal-fired boiler converts water into steam (Jagarnath 2019). Two boilers exist in coal-fired thermal plants: the pulverised coal-fired boiler and the fluidised bed combustion boiler. In India, 70% of electricity is produced by coal-fired power plants, and in China, thermal power generation contributes 73.4% to the total capacity (Bi et al. 2014; Mishra 2004). Coal generally has a low calorific value and a high ash content (Mishra 2004). The problem is that the electricity generated from these thermal power plants produces greenhouse gases, including nitrogen oxides $\left(N{O}_x\right)$ and carbon monoxide $\left(CO\right)$ (Bi et al. 2014). Therefore, the impact of power generation globally comprises high environmental and ecological risks.

From an engineering viewpoint, Haddadin et al. (2023) indicated that a balance could be reached by maintaining the current structure while increasing the use of renewable energy within power plants. This prospect will provide a dual benefit for a cleaner environment amid a few changes to the power system’s components and framework (Haddadin et al. 2023). Despite the need to shift from coal to other renewable energy sources, the generation and production of electricity in coal-fired thermal power plants will continue for at least another twelve years (Zima et al. 2022). Coal-fired power plants generate 38% of electrical power worldwide (Kitto 1996). For this reason, the focus should be on technological improvements, behavioural changes and the price or policy tools to induce these changes towards optimal efficiency (Lovins 2017).

One operational objective of the coal-fired boiler is to reach optimal operating efficiency while considering reliability and cost (Rambalee et al. 2011). Various technologies have been implemented to optimise and stabilise these thermal plants. These technologies include model predictive control (MPC) and advanced process control (APC), which have been introduced to ensure stable control of the boiler and to balance the burner management system (Rambalee et al. 2011). These technologies contribute to reducing harmful emissions ($i.e.,N{O}_x$ and $CO$).

With the rapid development of these technologies, a large quantity of data has been captured (Sleiti, Kapat, and Vesely 2022). The volume of data is increasing in all sectors, which results in more opportunities for improved efficiency (Narisco and Martins 2020). However, the structure and volume of the datasets are becoming more complex and newer methods are required to deal with these complexities (Sleiti, Kapat, and Vesely 2022). Most enterprises focus on machine learning (ML) to explore and operate these volumes of data (Narisco and Martins 2020). Machine learning is a subset of artificial intelligence (AI) (Shinde and Shah 2018). Ayodele (2010) stated that analysing machine learning algorithms is classified as computational learning theory. Subsequently, the learning is not considered consciousness learning but finding patterns in the data instead. Machine learning algorithms provide valuable insight into the difficulty of learning relationships in various environments (Ayodele 2010).

Machine learning techniques have been developed since the 1950s, and many advancements have been made within the last few decades. Prominent machine learning applications include the following industries: manufacturing (Wuest et al. 2016), retail (Jia, Zhao, and Tong 2013), finance (Goodell et al. 2021), and the energy sector (Nabavi et al. 2020). There has been an increase in academic papers on energy management and various new developments focusing specifically on machine learning and sustainability. However, Prashar (2017) stated that there is still a lack of knowledge on implementing these concepts for energy efficiency.

For this reason, this research study aims to investigate machine learning applications at coal-fired thermal power plants for improved energy efficiency by conducting a systematic literature review. To achieve the aim, the following objectives have been formulated:

1. To provide an overview of the fundamentals of machine learning theory followed by an overview of how these techniques are currently applied for energy efficiency.

2. To investigate which machine learning techniques are applied at coal-fired thermal power plants for improved energy efficiency.

3. To investigate how these machine learning techniques are applied at coal-fired thermal power plants for improved energy efficiency.

4. To identify the current literature limitations and to provide valuable future research recommendations.

The research paper is outlined as follows. First, a detailed literature study on machine learning algorithms is provided in Section 2, specifically on supervised, unsupervised, semi-supervised and reinforcement learning. Section 3 outlines the search protocol developed for the systematic literature review and discusses the results in Section 4. Conclusions and future recommendations based on the results obtained in this study are provided in Sections 5 and 6, respectively.

2. Literature review: the application of machine learning for energy efficiency improvements

Artificial intelligence (AI), in general, impacts the energy sector since it can assist in the development of clean, cheap, and reliable energy (Makala and Bakovic 2020). Furthermore, AI eliminates energy waste and reduces energy costs by improving the planning, operations, and control of various power systems (Makala and Bakovic 2020). Machine learning is a subset of AI closely related to statistics (Ongsulee 2017). First, an overview of the fundamentals of machine learning is provided, followed by a few case studies on how some of these algorithms are applied for improved energy efficiency in the industry.

2.1. An overview of machine learning

Machine learning techniques are divided into supervised, unsupervised, semi-supervised, and reinforcement learning (Bhatt 2019). Supervised learning develops an artificial system that maps the relationships or the patterns between the input and the output variables (Liu 2011). Based on these findings, the system predicts outcomes based on new information by referring to the relationships and patterns identified during the learning process. Classification of the input data is determined if the output takes a finite set of discrete values. In contrast, regression of the input data is identified if the output takes continuous values. With supervised learning, the input-output relationships are based on specific parameters. These unknown parameters must first be determined by training the algorithms (Liu 2011). The structure of supervised learning is demonstrated in Figure 1 (Liu 2011):

Figure 1. Supervised learning algorithm (Liu 2011).

With supervised learning, the training data need to be labelled (Liu 2011). Labelled data are already tagged with the specific output (Javatpoint 2022). The training sample is identified as $\left\{X_i,Y_i\right\}$. The variable $\left(X_i\right)$ is the system input, and $\left(Y_i\right)$is the system's output.

The indices (i) is defined as the index of the training sample. First, $\left(X_i\right)$is fed to the learning system, and an output $\left({\tilde{Y}}_i\right)$ $\stackrel{~}{}$ is generated. The output variable $\left({\tilde{Y}}_i\right)$ is compared to the labelling $\left(Y_i\right)$by an arbitrator. This arbitrator computes the difference between the $\left(Y_i\right)$ and $\left({\tilde{Y}}_i\right)$variables. The difference between the two variables, which is the error signal, is sent back to the learning system. The system evaluates the difference and adjusts the parameters accordingly to minimise the error signal, the difference between the $\left(Y_i\right)$ and$\left({\tilde{Y}}_i\right)$variables. The primary purpose of this learning process is to minimise the error signal of the training set, and to obtain the optimal set of learning system parameters (Liu 2011).

The objective of training the artificial system is to estimate the parameters based on known data, whereas testing aims to assess the predictions on unseen data. The data used in the training process are not used in the testing of the system. For this reason, it is not guaranteed that the system will perform the same in both instances. In addition, the training process is complex and lacks generalizability. Generalizability is defined when the system is trained to minimise training error and complexity (Liu 2011).

Supervised learning algorithms include, but are not limited to, linear classifiers such as the support vector machine, decision tree, and Bayesian networks (Ayodele 2010). Supervised learning can be used for discriminative pattern classification since the classes can be manipulated. However, with supervised learning, collecting and labelling unlabelled data is challenging. Various mistakes can occur, and it is time-consuming (Liu 2011). In the following sections, various machine learning algorithms are explained.

2.1.1. The support vector machine

A support vector machine is an algorithm that assigns labels to objects in the dataset (Boser, Guyon, and Vapnik 1992). The purpose is to find an optimal hyperplane within a specific N-dimensional space (Gandhi 2018). In this N-dimensional space, the data points are classified accordingly, as demonstrated in Figure 2 (Gandhi 2018).

Figure 2. The support vector machine: Potential hyperplane (Gandhi 2018).

The hyperplanes in N-dimensional space separate two classes from each other. The primary goal is to select the hyperplane with the maximum margin. A maximum margin is defined as the maximum distance between the two classes.

Gandhi (2018) stated that future data points could be classified with more confidence by maximising the distance. A hyperplane depends on the number of features. The hyperplane is a line if the support vector machine has two features. However, if the number of features is three, the hyperplane is two-dimensional. A support vector is a data point closest to the hyperplane. Subsequently, this vector influences the position and orientation of the hyperplane (Lovins 2017).

2.1.2. The decision tree

A decision tree is applied to solve classification and regression problems (Gupta 2017). An upside tree represents the decision-making process. The tree consists of internal nodes, which are split into branches. A decision is made at the end of each branch. A classification tree primarily aims to classify the data points according to the classes, whereas a regression tree predicts continuous values (Gupta 2017).

With a decision tree, all the features and attributes are considered in the decision-making process. The cost is linked to the accuracy of the prediction. For example, a function calculates how much accuracy each decision tree split will cost. The split with the least cost is selected. A cost function is typically used for both classification and regression problems. For regression, the following cost function is formulated: (1) $Regression:\sum (Y_i-{\tilde{Y}}_i{)}^2$(1) The average response of the training data of the group is considered the prediction for that group. The function is applied to all the data, and the cost is calculated accordingly. For classification, the following cost function is formulated: (2) $Classification:G=\sum _{k=1}^K\left(p_k\left(1-p_k\right)\right)$(2) The variable $\left(p_k\right)$ is the proportion within each class $k,k=1,2,\dots ,K$. A perfect class purity is defined when the $\left(p_k\right)$ value is either 1 or 0, meaning the group contains all inputs from a similar class. A minimum number of training inputs on each leaf can be defined for a decision tree. This will ensure that the tree stops splitting after reaching a specific number of splits. Subsequently, the parameter known as the maximum depth can also be set, which is defined as the length of the longest path from the tree's root to the leaf (Gupta 2017).

Other supervised machine learning techniques include the Bayesian network, Naive Bayes and the Gaussian Process (GP). The Bayesian network is a probabilistic model that mainly uses Bayesian inference to compute probabilities. The model consists of nodes and directed edges (Brownlee 2019). These models consider conditionally dependent and independent relationships between random data points. Like the rest of the algorithms discussed in this section, these models are also trained before predicting the probabilities of events.

The Naïve Bayes is defined as a probabilistic model based on the Bayes Theorem. The Naïve Bayes algorithm is mainly applied to solve classification problems. Bayes Theorem is defined as follows (Chauhan 2022): (3) $P\left(A|B\right)={\displaystyle \frac{P\left(B|A\right)P\left(A\right)}{P\left(B\right)}}$(3) where P(A|B) is the probability of A occurring, given B has already occurred. P(B|A) is the probability of B occurring given A has already occurred. P(A) is defined as the probability of A occurring, and P(B) is the probability of B occurring.

The Gaussian Process (GP) is designed specifically to solve regression and probabilistic classification problems. These processes are versatile, so various kernels can be specified, and the prediction interpolates the observations. However, these processes lose efficiency in high-dimensional spaces (Yi 2022).

Unlike supervised learning, unsupervised learning algorithms do not use labelled data (Ayodele 2010). Unsupervised learning is complex since the hidden patterns and relationships between the data points are unknown. Unsupervised learning algorithms primarily aim to identify hidden patterns and group similarities (Javatpoint 2022).

The unsupervised learning algorithm mainly addresses three problems (IBM Cloud Education 2020b): clustering, association, and dimensionality reduction. Clustering is defined as a method that clusters or groups objects in one specific group that has similarities (Javatpoint 2022). An association rule is a method that identifies the relationships between various variables in a dataset (Javatpoint 2022). The size of the datasets used influences the performance. However, overfitting is a common challenge when working with large data sets. Dimension reduction is applied to reduce the number of features. This algorithm also preserves the integrity of the data while reducing the variables to a reasonable size (IBM Cloud Education 2020b).

Unsupervised learning algorithms can solve complex problems in a considerable amount of time. However, the output might not be as accurate as supervised learning since the data is not labelled, and no result is known in advance (javatpoint 2022). One of the most popular unsupervised learning algorithms is neural networks (NN). This algorithm is discussed in the following section.

2.1.3. Neural networks

Neural networks (NN) are also known as artificial neural networks (ANN) or simulated neural networks (SNN) (IBM Cloud Education 2020a). These networks are a subset of machine learning at the centre of deep learning. Neural networks consist of node layers (input, hidden, and output layers). The structure of a deep neural network is presented in Figure 3 (IBM Cloud Education 2020a).

Figure 3. The deep neural networks (IBM Cloud Education 2020a).

Each node or artificial neuron consists of a weight and threshold. As stated in Section 2.1, each algorithm must be trained to improve accuracy over time. Training of the algorithm contributes towards the accuracy of the prediction. Once properly trained, the algorithm can be employed as a powerful machine learning tool, specifically for image and speech recognition. A node is similar to a linear regression model, consisting mainly of data, weights, a threshold value, and an output. A node’s equation is defined as follows (IBM Cloud Education 2020a): (4) $\sum _{i=1}^m{w}_i{x}_i+bias=w_1{x}_1+w_2{x}_2+w_3{x}_3+\dots +w_m{x}_m+bias$(4)

Weights $\left(w_i\right)$ are assigned to the node when the input layer is determined. The weight depends on the importance of the provided variable (IBM Cloud Education 2020a). An activation function is used to determine the output. An activation function is defined as follows (IBM Cloud Education 2020a): (5) $Output=f\left(x\right)=\{\begin{array}{c}1if\sum _i{w}_i{x}_i+b\ge 0\\ 0if\sum _i{w}_i{x}_i+b<0\end{array}$(5) If the predicted output of a specific node is above the threshold value, the node is activated, and data is sent to the following layer of the entire network. Therefore, the output of one node becomes the input of another. A feedforward neural network is defined when data is passed from one node to another in the network (IBM Cloud Education 2020a).

There exist various types of artificial neural networks. Only the most common types are discussed in this paper. Figure 4 depicts the most basic neural network as demonstrated by Frank Rosenblatt in 1958 (Rosenblatt 1958):

Figure 4. The neural networks (Rosenblatt 1958).

Feedforward neural networks are also referred to as multi-layer perceptions (MLPs). They are comprised of sigmoid neurons and not perceptions since most real-world problems are non-linear. Convolutional neural networks (CNN) typically use principles from linear algebra, specifically matrix multiplication, to identify specific relationships in the data. Recurrent neural networks (RNN) consist of various feedback loops. Time-series data are used to make precise predictions (Rosenblatt 1958).

Johnson (2022) stated that one of the most common ways to train neural networks is to use the backpropagation method. This method is applied by tuning the neural network weights by considering the error rate obtained from the previous epoch. As explained in Section 2.1, the primary purpose of training any machine learning algorithm is to reduce the error rate to deliver a more reliable model for accurate predictions. This approach is mostly followed in training neural networks since it is fast and easy to programme. There are no parameters to tune except the number of inputs, and it is considered flexible (Johnson 2022).

A study conducted by Smrekar et al. (2009) employed machine learning by developing an artificial neural network (ANN) model that predicts steam properties from a coal-fired boiler. The authors considered three input parameters: steam pressure, temperature, and flow. The models developed in this study yielded high accuracy on both the training and testing data. Furthermore, the study demonstrated the usefulness of ANN modelling for coal-fired boilers. The authors stated that some boiler processes, such as pulverises and combustion, are challenging to develop physical models since the relations between these variables are not always known. For this reason, ANN models are applicable since these types of models can consider most things known and unknown (Smrekar et al. 2009).

Other unsupervised machine learning techniques include the K-Nearest Neighbor and the K-means clustering. The K-Nearest Neighbor is applied to solve both classification and regression problems. The K-Nearest Neighbor assumes that similar characteristics exist between data points near each other. The euclidean distance is used to calculate the distance between the data points (Harrison 2018).

The K-means clustering technique divides data points into K groups, where K is defined as the number of clusters (IBM Cloud Education 2020b). The number of clusters is based on the distance from each group's centroid. All points allocated near the centroid are clustered into one category. A small K-value indicates large groupings with less granularity, and a large K-value means smaller groupings with more granularity.

Semi-supervised learning is defined when the data is only half labelled (IBM Cloud Education 2020b). Subsequently, unsupervised and semi-supervised learning is mainly applied to prevent data from being incorrectly labelled.

2.1.4. Reinforcement learning

Supervised and reinforcement learning utilises a function that maps the dependency between the input and output data (Song et al. 2013). However, supervised learning provides a correct set of actions to obtain a specific output, whereas reinforcement learning uses rewards and punishments to obtain a particular result (Bhatt 2019).

Unsupervised learning and reinforcement learning's objectives are different. Unsupervised learning recognises the patterns between data points, whereas reinforcement learning identifies an action model. This action model aims to maximise the total cumulative rewards of the system. Figure 5 provides an overview of the action-reward feedback loop of reinforcement learning (Bhatt 2019):

Figure 5. Reinforcement learning algorithm (Bhatt 2019).

The state $\left(S_t\right)$ is defined as the agent's current state. The reward variable $\left(R_t\right)$ indicates the feedback from the environment, and a policy refers to the method used to map the agent's actions (Bhatt 2019). The environment is defined as the space the agent operates in, and the value $\left(R_{t+1}\right)$ is defined as the future reward the agents may receive if a specific action $\left(A_t\right)$ is followed. The primary objective is to build an optimal policy while maximising the reward. In reinforcement learning, exploration and exploitation are prevalent definitions. A decision must be made where both are balanced (Bhatt 2019).

2.1.5. Errors and evaluation criteria

The performance of a machine learning model is analysed based on the prediction error against the validation dataset. This study presents and discusses the coefficient of determination $\left(R^2\right)$, root-mean-square error (RMSE), normalised RMSE, mean absolute error (MAE) and mean absolute percentage error (MAPE). The values obtained from these performance metrics provide information on the machine learning model's robustness and effectiveness (Muhammad Ashraf et al. 2020).

Assume $\left(n\right)$ presents the sample size used as input, $\left(y_i\right)$and $\left({\hat{y}}_i\right)$ denote the predicted values and the actual values, respectively. The maximum and minimum value of $\left(y_i\right)$ are denoted by $\left(y_{max}\right)$ and $\left(y_{min}\right)$, respectively. The coefficient of determination $\left(R^2\right)$ is given in Equation (6). (6) $R^2=1-{\displaystyle \frac{S_{RESIDUAL}}{S_{TOTAL}}=1-{\displaystyle \frac{{\sum }_{i=1}^n{\left(y_i-{\hat{y}}_i\right)}^2}{{\sum }_{i=1}^n{\left(y_i-{\displaystyle \frac{1}{n}{\sum }_{i=1}^n{y}_i}\right)}^2}}}$(6)

Nagelkerke (1991) indicated that the coefficient of determination estimates the percentage of variance explained in a given dataset. The root-mean-square error (RMSE) and the normalised RMSE are estimated as illustrated in Equations (7–8) (7) $RMSE=\sqrt{{\displaystyle \frac{1}{n}\sum _{i=1}^n{\left(y_i-{\hat{y}}_i\right)}^2}}$(7) (8) $NRMSE={\displaystyle \frac{RMSE}{y_{max}-y_{min}}\times 100\text{\%}}$(8)

The root-mean-square error (RMSE) estimates the standard deviation of the residuals between the predicted values and the actual values of an observation (Otto 2019). Furthermore, the normalised RMSE (NRMSE) enables comparing various models with different scales (Otto 2019). The mean absolute error (MAE) and mean absolute percentage error (MAPE) are illustrated in Equations (9–10) (9) $MAE={\displaystyle \frac{1}{n}\sum _{i=1}^n\left|y_i-{\hat{y}}_i\right|}$(9) (10) $MAPE={\displaystyle \frac{1}{n}\sum _{i=1}^n\left|{\displaystyle \frac{y_i-{\hat{y}}_i}{y_i}}\right|\times 100\text{\%}}$(10) The mean absolute error (MAE) presents the error obtained between the predicted values and the real values of the specific dataset (Qi et al. 2020). Furthermore, the mean absolute percentage error (MAPE) estimates the average magnitude of the error between the predicted value and the actual values of a specific dataset (De Myttenaere et al. 2016).

Baressi Šegota et al. (2020) stated that the K-fold cross-validation is part of the training process of a machine learning model, specifically when the data sets used are small. First, the data is split into k splits, where (k−1) subsets of the dataset are used for training, and the remaining subset is used for testing. This process is repeated k times until all the splits have been used as the testing split precisely once. Subsequently, this allows the entire dataset to be used as a testing set systematically and provides more information on the model's performance (Baressi Šegota et al. 2020).

2.2. The application of machine learning for improved energy efficiency

Machine learning applications also include, but are not limited to, renewable energy predictions, fault detection and power quality monitoring. Renewable energy resources, including wind power, hydropower, and photovoltaic energy, have gained attention over the last few years and are considered alternative energy supply (Sogabe et al. 2018). The problem encountered with renewable energy resources is often a mismatch between the electricity supply and demand, resulting in instability and limited power output (Sogabe et al. 2018). Virtualitics (2021) added that the power demand and generation must always be equivalent to prevent blackouts due to inefficient energy.

However, machine learning algorithms can predict the electricity required for any renewable source for a specific period to satisfy demand. Virtualitics (2021) argued that these predictions and evaluations will contribute towards reducing harmful emissions by encouraging the use of reliable renewable energy sources as an alternative to coal or other non-renewable energy. Sogabe et al. (2018) indicated that there is a need for a practical approach to address these types of issues. Khan and Khan (2013) proposed smart grids due to their ability for intelligent, robust, and functional power grids. Sogabe et al. (2018) applied two deep reinforcement learning techniques for smart grid optimisation. These techniques captured the supply and demand feature and selected the appropriate behaviour, maximising the reward of the deep reinforcement learning model (Sogabe et al. 2018). The beforementioned studies provide a general overview of how these algorithms can assist with better decision-making.

Despite using renewable energy sources, clean coal power plants are still one of the dominant choices for power generation by developing energy-efficient operations and ensuring carbon capture and storage (Haddadin et al. 2023). Haddad and Mohamed (2022) stated that, with the newest advancements in clean coal technologies, it is essential to gain a fundamental and comprehensive understanding of these modelling philosophies. For this reason, this study focuses on machine learning applications at coal-fired thermal power plants for improved energy efficiency.

One of the leading causes of failure in a steam power plant is defined as boiler waterwall tube leakage (Khalid et al. 2020). In 2021, Khalid, Hwang, and Kim (2021) stated that the primary reason for failure in thermal power plants is equipment, specifically boiler and turbine faults (Khalid, Hwang, and Kim 2021). Based on these two studies, the authors indicated that developing efficient fault-detection processes in power plants is required. These intelligent tube leak-detection systems can improve the efficiency and reliability of modern power plants (Khalid et al. 2020; Khalid, Hwang, and Kim 2021). Khalid, Hwang, and Kim (2021) proposed a novel machine learning-based optimal sensor selection to analyse the turbine and boiler faults in the power plant. The results of this study demonstrated that the machine learning models’ computational efficiency was enhanced by reducing the number of sensors for both scenarios. Furthermore, the performance of these models was improved by up to 97% and 92.60% in the water wall tube leakage and the turbine motor fault, respectively (Khalid, Hwang, and Kim 2021).

Alvarez Quiñones, Lozano-Moncada, and Bravo Montenegro (2023) proposed a methodology to schedule predictive maintenance of distribution transformers using machine learning. The authors focused explicitly on classification predictive modelling to obtain the minimal number of distribution transformers prone to fail. In conclusion, the machine learning methodology was implemented and saved 13% in corrective maintenance expenses in one specific year (Alvarez Quiñones, Lozano-Moncada, and Bravo Montenegro 2023).

In the study conducted by Haddadin et al. (2023), the Elman neural network (ENN) and the generalised regression neural network (GRNN) were proposed for a cleaner coal-fired supercritical unit. These models obtained more accurate results than previously published models. Subsequently, the authors predicted the supercritical pressure, the production, and the steam temperature. It is important to note that these variables have been selected since they represent energy efficiency and cleaner production. The Mean Square Error was used as a performance metric to evaluate the models’ effectiveness and performance (Haddadin et al. 2023).

From these case studies, it is evident that the application of machine learning is extended to solve a wide range of problems in various areas, including renewable energy predictions, fault detection and optimal power generation in coal-fired thermal power plants. Furthermore, machine learning provides the ability to make more accurate decisions in real-time and deliver sustainable solutions (Ghosh 2017). To further extend the machine learning analysis, an in-depth investigation is conducted on the type of machine learning algorithms and how these algorithms were applied in coal-fired thermal power plants for improved energy efficiency from 2012 - 2022. First, the research method of the systematic literature review is outlined, followed by a discussion of the results.

3. A systematic literature review: the review protocol

The number of published papers in various research fields exponentially increases, resulting in information overload. Pejić-Bach and Cerpa (2019) stated that there is a gap between the number of published articles and the time to review these studies. A systematic literature review addresses this gap by providing insight into only the relevant papers and their scientific development (Pejić-Bach and Cerpa 2019).

A systematic literature review is considered evidence-based research and originated in the medical industry (Kitchenham et al. 2009). Kitchenham et al. (2009) defined the word evidence as a synthesis of the best quality studies in the academic literature of a particular topic or research field. A systematic literature review is different from a formal literature review since a methodologically rigorous method is followed to obtain research results (Kitchenham et al. 2009).

In this paper, a systematic literature review is conducted to investigate machine learning applications at coal-fired thermal power plants for improved energy efficiency. The systematic literature review conducted in this study is based on the guidelines proposed by Kitchenham et al. (2009) and Kumar and Garg (2020). Kitchenham et al. (2009) conducted a systematic literature review to identify the status of Evidence-based Software Engineering (EBSE) papers since 2004, and Kumar and Garg (2020) conducted a systematic literature review on context-based sentiment analysis in social multimedia.

Kitchenham et al. (2009) method consists of seven stages, including the formulation of research questions, search process, inclusion and exclusion criteria, quality assessment, data collection, data analysis and deviations from the protocol. The review process by Kumar and Garg (2020) was divided into six stages: formulating the research questions, search strategy, study selection, quality assessment, data extraction and data synthesis. The search protocol followed in this paper consists of five stages: Formulating the research questions, the search process, study selection: Inclusion and exclusion criteria, quality assessment, data extraction, and data synthesis.

3.1. Research question formulation

The following research questions were formulated to collect and collate research evidence from existing studies:

1. Research Question 1: Which machine learning algorithms were applied at coal-fired thermal power plants for improved energy efficiency?

2. Research Question 2: How were these machine learning algorithms applied at coal-fired thermal power plants for improved energy efficiency?

3.2. Search process

The search process consists of defining keywords in locating only the relevant studies. Scientific papers that contain all three keywords were selected:

1. Machine learning

2. Energy efficiency

3. Coal-fired thermal power plant

A manual search was conducted, and only published scientific journal articles were selected. The following databases were used to identify the relevant studies:

1. Science Direct

2. Scopus

3. Web of Science

4. IEEE Explore

5. Emerald Insight

3.3. Study selection: inclusion and exclusion criteria

The following studies were included in the search process:

1. Only published journal articles were included in the search process.

2. Journal articles containing the following three keywords: Machine learning, energy efficiency and coal-fired thermal power plant. These studies may include any review study or studies that refer to other studies containing the relevant keywords.

3. Journal articles that were published between 2012 and 2022.

The following studies were excluded from the search process:

1. Non-English articles

3.4. Quality assessment process

For the quality assessment, Kitchenham et al. (2009) followed the database of abstracts of reviews of effects (DARE) criterion outlined by York University. This criterion is explicitly defined for studies that evaluate other studies that conducted a systematic literature review. The systematic literature review conducted in this study identified all types of studies. For this reason, quality assessment criteria are used as outlined by Albliwi et al. (2014) and Mangaroo-Pillay and Coetzee (2022). This assessment consists of four phases: Identification, screening, eligibility, and inclusion. Each phase is described in Table 1.

Table 1. The quality assessment process.

Phases	Description of each phase
Identification	The databases were searched using the defined keywords in each study's title, abstract and keywords. The following results were obtained: 1. Science Direct = 2 2. Scopus = 11 3. Web of Science = 0 4. IEEE Explore = 1 5. Emerald Insight = 0 Total = 14
Screening	The following results were obtained by screening the studies to ensure that each study adheres to the inclusion and exclusion criteria, as outlined in Section 3.3. 1. Science Direct = 2 2. Scopus = 11 3. Web of Science = 0 4. IEEE Explore = 1 5. Emerald Insight = 0 Total = 14
Eligibility	All duplicates were removed after screening each study to ensure the studies met the inclusion and exclusion criteria. No duplicates were identified. 1. Science Direct = 2 2. Scopus = 11 3. Web of Science = 0 4. IEEE Explore = 1 5. Emerald Insight = 0 Total = 14
Inclusion	A full-text assessment was done on all 14 studies to investigate machine learning applications at coal-fired thermal power plants for improved energy efficiency. The results are discussed in Section 4.

3.5. Data extraction

The data extracted from each study were the following:

1. The fundamentals of each study (Table 2).

2. The type of machine learning algorithm applied (Table 2).

3. The year of publication (Table 2).

4. The databases (Table 2).

Table 2. Systematic literature review results.

#	Title of the study	Type of machine learning model applied	Year	Database	Reference
1	Digital twin in the energy industry: Proposed a robust digital twin for power plants and other complex capital-intensive large engineering systems.	A deep reinforcement learning (DLR) model	2022	Science Direct	Sleiti, Kapat, and Vesely (2022)
2	Energy digital twin technology for industrial energy management: Classification, challenges, and future.	Review paper: The existing literature on energy digital twin technology was investigated in this study. The authors highlighted the advantages, including improvements in energy management and optimisation, improved maintenance and servicing, and enhanced renewable energy. This study proposes an original multi-dimensional digital twin with guidelines demonstrating how these technologies can be applied to reduce the carbon footprint	2022	Science Direct	Yu et al. (2022)
3	Thermal calculations of a natural circulation power boiler operating under a wide range of loads.	Other – A methodology was developed specifically for the thermal flow of a natural circulation power boiler and contributed towards the power boiler's design process. Various in-house methods were formulated and used in conjunction with the literature. This method is developed to compute the heat transfer in the combustion gas duct's wall superheaters and hanger tubes. Following these calculations, a computer programme was developed. Various other CFD computations were performed that focused explicitly on the boiler furnace chamber. The results were compared to the data provided by the manufacturer. Only slight differences were observed, and the models were verified.	2022	Scopus	Zima et al. (2022)
4	A Review on the Adoption of AI, BC, and IoT in Sustainability Research	Review paper: Artificial intelligence (AI), blockchain (BC), and the internet of Things (IoT) are the factors influencing the advancement of sustainability the most. This study aimed to investigate how digital transformation drives natural and human systems by conducting a bibliometric analysis. The authors assessed 960 studies, concluding that most advancements were made in the following sectors: Smart cities, energy systems and supply chain management. Artificial intelligence contributed towards these advancements by assisting with scheduling, predicting, and monitoring energy systems within smart cities. In conclusion, these authors stated that there is an increase in the technological integration of AI and IoT, specifically in the smart city and energy sectors.	2022	Scopus	Wu et al. (2022)
5	The Implementation of K-Means Clustering in Coal-fired Power Plant Performance.	K-Means clustering	2022	Scopus	Wahyuningtyas and Asrol (2022)
6	A systematic review of big data in energy analytics using energy computing techniques	Review paper: Energy computing techniques in the energy-orientated domain were investigated in this study. The authors indicated an increase in energy analytics due to the rapid increase in data. In conclusion, this study highlighted that the thermal power plant, smart grid, and power systems datasets were extensively utilised, and the artificial neural network (ANN) was mainly applied.	2022	Scopus	Dhanalakshmi and Ayyanathan (2022)
7	Repowering Industrial Combined Heat and Power Units: a contribution to Cleaner Energy Production	Other – This study is classified as a gas turbine-based repowering study. Repowering industrial units results in a reduction of greenhouse gas emissions. This study specifically focused on fuel and carbon monoxide reduction. In conclusion, two conservative repowering options were identified, and both indicated feasible economics with a simple payback period.	2022	Scopus	Variny and Kšiňanová (2022)
8	Improvement of marine steam turbine conventional exergy analysis by neural network application	An Artificial neural network – A multi-layered perceptron (MLP)	2020	Scopus	Baressi Šegota et al. (2020)
9	Optimisation of a 660 MWE supercritical power plant performance: A case of industry 4.0 in the data-driven operational management. Part 2: power generation	The following machine learning methods were developed in this study: 1. Artificial neural network (ANN): 2. Multiple linear regression (MLR): 3. The least-square vector machine (LSSVM)	2020	Scopus	Muhammad Ashraf et al. (2020)
10	Artificial neural network and its applications in the energy sector – An overview	Review paper: This study aimed to summarise artificial intelligence methods’ significance in the energy sector. The authors indicated that artificial neural networks are effective methods designed to solve complex problems within modelling, control, and optimisation.	2020	Scopus	Babatunde (2020)
11	Data-driven framework for boiler performance monitoring	Other – This study aimed to propose a framework for monitoring a steam boiler's real-time operations. The boiler's actual efficiency was compared and monitored with the expected efficiency. The performance of the boiler is measured using a statistical process control chart. The framework was verified using fluidised bed boiler and corner-fired boiler historical data. In conclusion, the strongest correlations between the variables and boiler efficiency were consistent in both cases.	2016	Scopus	Nikula, Ruusunen, and Leiviskä (2016)
12	Neuro-genetic model for crude oil price prediction while considering the impact of uncertainties	An artificial neural network model combined with the genetic algorithm (neuro-genetic)	2016	Scopus	Chiroma and Abdulkareem (2016)
13	A methodology for energy savings verification in the industry with the application for a CHP (combined heat and power) plant	An Artificial neural network (ANN)	2015	Scopus	Rossi and Velázquez (2015)
14	Control Strategy for Denitrification Efficiency of a Coal-Fired Power Plant Based on Deep Reinforcement Learning	A deep reinforcement learning (DLR) model	2020	IEEE Explore	Fu et al. (2020)

One researcher extracted the data, and the other checked the extraction. All three researchers evaluated the selected studies to ensure that the studies adhered to the quality assessment process.

3.6. Data analysis

Each study identified is summarised in Table 2 to show:

1. The title of each study

2. The type of machine learning algorithm applied in coal-fired thermal power plants for energy efficiency.

3. Each database the study was published in.

4. The year of publication.

5. Reference.

4. Discussion of results

In this section, the results are discussed.

4.1. Search results

As indicated in Section 3, the review results are collected by the following information: Title, the type of machine learning model applied, published year, database, and reference. The studies identified were divided into three sections: Those that applied machine learning models, review papers and others. Others are defined as papers that did not apply machine learning models; however, the authors of these studies referred to other studies that contained the keywords used in this systematic literature review. Therefore, these studies were included in the search. All reviews and other studies are explained. The studies that applied machine learning models are further discussed in Table 3. For this reason, only an overview of the model type is provided in Table 2.

Table 3. Describes how machine learning models were applied in each study.

#	Title of the study	An overview of each machine learning application	Contribution(s) and Future Research
1	Digital twin in energy industry: Proposed a robust digital twin for power plants and other complex capital-intensive large engineering systems	This study did not specifically apply to machine learning. However, deep learning is used to assist the authors in developing a database model. To achieve high reliability while considering cost, a digital twin (DT) architecture is required for future power plants. The paper consists of three sections: The authors provide an overview of digital twins and research regarding energy applications. The design requirements for the digital twin are established by considering physics-based formulation, sensor data analysis, statistical analysis, and the system Genome. Valuable future recommendations are provided. The authors highlighted that a data-driven approach alone is not sufficient for achieving improved reliability. They recommended that a low-order physics model should operate with the updated system parameters to enhance the results of the data-driven approach. The database model is designed to use ADL to detect and identify the differences between the dynamic systems model (DSM) and the digital twin (DT). The sensor failures are identified using deep learning and are sent back to the sensor database. The sensor database is updated accordingly. In conclusion, this study demonstrated the discrepancies between the dynamic system models (DSM), anomaly detection, and deep learning (ADL). Using operational data, the vector autoregressive model was applied for anomaly detection in utility gas turbines.	The contributions of this study include the following: A comprehensive overview of digital twins subject to the application, frameworks, and architectures for power plants. Research related to digital twins and how these can contribute towards energy savings applications. The authors also proposed a robust digital twin for power plants. Future research includes the following: Any digital twin requires actual data and a description representing a holistic view of the system, focusing explicitly on each part and connection. The authors suggested that data-driven algorithms capable of predicting dynamic system behaviour are still lacking in the literature. A data-driven model alone is not sufficient. For this reason, the authors proposed that a physics-based model must be integrated with system parameters to improve the interpretation of the results from the data-driven processes. The authors noted that research related to cyber security with digital twins had not been reported in the literature, which can be a future research study.
2	The Implementation of K-Means Clustering in Coal-fired Power Plant Performance	The k-Means Clustering algorithm is developed for coal-fired power plant performance. Two attributes are required for power plant performance: Equivalent Availability Factor (EAF) and Equivalent Outage Rate (EFOR) The data set used in this study is from 2020, and the data performance indices in these investigations are the abovementioned attributes. The authors obtained the data set from the Generation Availability Information System (GAIS) database. The authors pre-processed the data to ensure data integrity. A number of 576 numerical datasets of the two attributes were pre-processed. The relationship between the EAF and the EFOR was investigated to determine the characteristics of the power plant. The analysis concluded a negative correlation between the two attributes, and the authors did a further investigation. The features of the input data determined the number of clusters, and the Elbow method was applied to analyse the total number of clusters used explicitly for power plant clustering. After an in-depth investigation, 4 clusters were selected for optimal power plant performance. This number of clusters produced each cluster's highest EAF value and the lowest EFOR value In conclusion, the higher the EAF value and the lower the EFOR value, the better the plant performs. For this reason, the authors evaluated these values and made conclusions accordingly. Applying this machine learning technique, the results indicated which plant needs to be prioritised for maintenance.	The contributions of this study include the following: The authors demonstrated the ability of machine learning to analyse current power plant conditions successfully. Future research includes the following: A real dataset is required to control current conditions to improve the decision-making process. Furthermore, the authors illustrated that a decision-support system may facilitate the decision-maker in making faster and more reliability decisions.
3	Improvement of marine steam turbine conventional exergy analysis by neural network application.	Neural networks were applied to improve the steam turbine's conventional exergy. Subsequently, a multi-layered perceptron (MLP) model is developed in Python programming language using the Scikit-Learn library. The data set comprised 125 points for the ambient temperature, pressure, steam mass flow rate, and exergy destruction and efficiency values for the high-pressure cylinder, low-pressure cylinder, and the entire turbine. The input variables were the steam temperature, pressure, and mass flow rate. The weights of the inter-neuron connection and the values of the hidden layers are randomly defined. The output values are compared to the input values by focusing explicitly on exergy destruction or efficiency. Each data point has 23 input values and approximately six output values. The backward propagation algorithm is used to train the neural network, and this method is specifically used to adjust the weights based on the gradient of the error. The training process is repeated until a very low error is obtained for the training model. It is important to note that two models were trained for each possible output: exergy destruction and exergy efficiency. For this reason, this study had a total of 72 final models. One of the purposes of training the model is to identify the optimal hyperparameter values. The hyperparameters selected were the activation function of the hidden layer neurons, the total number of hidden layers, the total number of neurons per hidden layer, and the solver, which estimates the weights, the learning rate and the L2 regularisation parameter. In this study, Grid Search was used using the GridSearchCV Function in Python. The training of the model took place on a supercomputer. The coefficient of correlation ($R^2$) and the mean absolute error (MAE) were used to determine the model's performance. The K-fold cross-validation method was applied with ten folds, and the data set was randomly split into training and testing data.	The contributions of this study include the following: Developing a neural network to improve the marine steam turbine conventional exergy analysis. This study demonstrated the usefulness of Artificial intelligence and that these algorithms could be applied to calculate the exergy destruction and exergy efficiency of a marine steam turbine. Future research includes the following: The authors suggested that a future study can be conducted to determine the optimal turbine operating points, precisely steam temperature, pressure, and mass flow rate. These optimal operating points may be integrated with the MLP to analyse exergy parameter prediction focusing on the entire turbine. This study's same approach and technique can be applied to other marine steam turbines. To improve the prediction of the turbine exergy analysis parameter, the authors suggested that a study be conducted where extensive measurements over a long period are collected on the turbine and the turbine's cylinders. Using these measurements, a more accurate prediction can be made based on the operating period.
4	Optimisation of a 660 MWE supercritical power plant performance: A case of industry 4.0 in the data-driven operational management. Part 2: power generation.	This study considers 24 thermo-electric parameters of a specific power plant to model the generator power under different power generation scenarios. These operating parameters were obtained from the power plant's turbine, boiler, and generator sides. The data set consisted of 1900 data points. The following machine learning methods were developed in this study: Artificial neural network (ANN): The feedforward backpropagation algorithm is developed in this study. The authors used the gradient descent with momentum as a training function. The tangent hyperbolic and purelin were employed as transfer functions in both the hidden and output layer of the ANN model. Multiple linear regression (MLR): The least square vector machine (LSSVM) was used to develop this model. First, the difference (referred to as residuals) between the actual and observed values was calculated. Second, the sum of the square of residuals is minimised to obtain the best fit i.e. least squares minimisation was employed. The least-square vector machine (LSSVM): This model was trained based on the structure risk minimisation (SRM) principle. The relationship between the input and output variables is determined by using a Gaussian kernel function. All variables were standardised for model training. The authors used the Bayesian optimiser and the expected improvement per second in conjunction with the acquisition function to optimise the model under 30 epochs. After training the MLR, ANN and LSSVM models, the models were validated using new (unseen) data. The unseen data were defined as generator power data and were not included in the training dataset. The coefficient of determination, mean absolute percentage error (MAE), root-mean-square error (RMSE), normalised RMSE (NRMSE), and the mean absolute percentage error (MAPE) were estimated to analyse the model's robustness and effectiveness. The results demonstrated that the LSSVM outperformed the other two models in predicting the plant's generator power. This model was further used to predict the generator power of the plant while considering numerous operating strategies.	The contributions of this study include the following: The ability of a least-square vector machine (LSSVM) model is demonstrated in this study. This model outperformed the multiple linear regression (MLR) models and the artificial neural network (ANN). This model's reliable performance in modelling generator power has been highlighted in this study. In this study, the power plant's operation control has been improved by employing data-driven optimisation strategies.
5	Neuro-genetic model for crude oil price prediction while considering the impact of uncertainties	The authors proposed a neural network model combined with the genetic algorithm (neuro-genetic) to predict the crude oil price while considering uncertainty. The parameters optimised are the neural network's bias, weights, and hidden layer nodes. The values for the input neurons, the output neurons, the activation function, the weights, and the hidden layers were defined prior to training the model. The model was trained by splitting the dataset into 80% training and 20% testing. The optimal parameter values were obtained through trial and error, and the sensitivity of the neuro-genetic model was analysed. The models were trained on two datasets obtained from the Energy Information Administration of the US Department of Energy. The first dataset used in this study was obtained from May 1987 to July 1991, excluding uncertainties. The second data set, including uncertainties, was obtained from August 1990 to February 1991. The proposed method was implemented using Matlab 2012b. The performance of the neuro-genetic model was investigated and compared to the neural network and support vector machine. It was concluded that the neuro-genetic model performance is better due to accuracy and CPU processing time.	The contributions of this study include the following: The authors proposed an artificial intelligence approach to predict the crude oil price while considering uncertainties. The neuro-genetic model developed outperformed the neural network (NN) and the support vector machine (SVM). Considering uncertainty in decision-making provides a more realistic and practical application of crude oil prediction. Future research includes the following: Advancing and updating the current model to consider the prediction of crude oil in different prediction horizons. The framework may be modified to be applied to optimise nuclear data libraries and nitride fuel.
6	A methodology for energy savings verification in the industry with the application for a CHP (combined heat and power) plant	A methodology to assess the savings achieved from developing and implementing energy conservation measures (ECM) in specific industrial plants was developed. An artificial neural network was developed in this study. Historical data were obtained from the plant's distributed control system and are pre-processed to ensure data integrity. The data were cleaned by eliminating invalid and non-representative data points. The applicable variables were defined, and the non-applicable variables were removed from the dataset. The authors indicated that the ANN is famous for detecting non-linear patterns and controlling many industrial processes and power plants. The ANN is also capable of replicating these processes with satisfactory levels of accuracy. The data samples were determined for training, followed by the selection of input variables. The ANN was trained using the back error propagation algorithm. The model's performance was analysed using the coefficient of determination, the root mean square error (RMSE), and the mean absolute error (MAE). Acceptable model performance was confirmed.	The contributions of this study include the following: The authors proposed a methodology to evaluate the savings from implementing energy conservation measures in industrial plants. Future research includes the following: The authors suggested evaluating other techniques, excluding the ANN model, including the feature selection method. These methods could improve the screening stage of the dataset. It is also important to note that feature selection is essential when working with large datasets and ranking all variables’ relevance to the target variable.
7	Control Strategy for Denitrification Efficiency of a Coal-Fired Power Plant Based on Deep Reinforcement Learning	In this study, a 1000MW unit is used as a case study. A deep reinforcement learning (DLR) model is developed with a long short-term memory model, the Asynchronous Advantage Actor-Critic algorithm (A3C). The data used in this study was obtained from the plant's Distributed Control System (DCS). The authors used 10000 sets of data from a 1000MW coal-fired boiler. Variables, including, but not limited to, spray ammonia mass flow, Inlet mass concentration, boiler load, inlet mass and inlet flue gas temperature, were identified. The data was pre-processed, followed by using this data as input for the Long short-term memory neural network (LSTM). Subsequently, this LSTM model was used for the DRL's environment. The DLR method used the state value output by the environment as an input to produce a new updated action value. By conducting several iterations, a new control strategy of SCR was obtained. The Back Propagation algorithm was used to train the LSTM. For the training, six main variables were defined as input variables; the model consisted of 3 LSTM layers and 64 nodes per layer. The model was optimised using the Adam optimiser algorithm, and the data were split into 20% testing and 80% training data. Two performance metrics were used: the root mean square error (RMSE) and the R-squared. The accuracy of the efficiency calculated was 91.7%. In conclusion, it was shown that the approach is feasible and universally applicable.	The contributions of this study include the following: One of the primary contributions is that the approach proposed in this study demonstrates the ability to be feasible and universally applicable in industry. Future research includes the following: The authors indicated that the following should be further investigated: the control method of the combustion process, the system cost of coal-fired power plants and the denitrification process. These elements should be investigated to improve the stability and reliability of control strategies.

Table 2 provides an overview of each study identified in the search by referring to the following: Title, type of machine learning model applied in each study, published year, database, and the reference. According to the search protocol developed, the studies were further analysed by addressing the two research questions.

4.1.1. Research question 1: which machine learning algorithms were applied in coal-fired thermal power plants for improved energy efficiency?

Fourteen studies contained all three keywords: Machine learning, energy efficiency and coal-fired thermal power plant. Two studies were identified in Science Direct, 11 in Scopus and one in IEEE Xplore. Seven of these studies were published in 2022, four in 2020, two in 2016 and one in 2015. These results are summarised in Figure 6.

Figure 6. Search results: Publications per year and database.

Figure 6 demonstrates that most studies were identified in Scopus, and zero studies were identified in Web of Science and Emerald Insight. Furthermore, most studies were published recently, in 2022, and fewer studies were published before 2015. These results indicate the increased attention machine learning applications have gained over the last decade at coal-fired thermal power plants for improved energy efficiency. Various machine learning algorithms were developed and applied in the 14 studies identified. Figure 7 provides an outline of these types of algorithms.

Figure 7. The type of machine learning algorithms applied in each study.

The types of machine learning algorithms developed in these studies were identified as follows:

• K-Means clustering

• The artificial neural networks (ANN)

• Multiple linear regression (MLR)

• The least-square vector machine (LSSVM)

• Deep reinforcement learning (DRL) and deep learning.

Muhammad Ashraf et al. (2020) applied three different machine learning algorithms. For this reason, 16 algorithms are identified in Figure 7. These results showed that artificial neural networks were the most popular algorithm to solve the identified problems. The multi-layered perceptron (MLP) was mainly developed and applied within these.

Four studies conducted a review: Yu et al. (2022), Wu et al. (2022), Dhanalakshmi and Ayyanathan (2022), and Babatunde (2020). These reviews did not particularly apply or develop machine learning models but provided an overview of these algorithms. The ‘other’ studies identified in Figure 7 did not apply or develop machine learning models: Zima et al. (2022), Variny and Kšiňanová (2022), and (Nikula, Ruusunen, and Leiviskä 2016). However, the authors of these studies referred to other studies that contained the keywords used in this systematic literature review. Therefore, these studies were included in the search. From the studies identified, it is also important to note that no review, like the one conducted in this study, was identified. Research question two is addressed in the next section.

4.1.2. Research question 2: how were these machine learning algorithms applied in coal-fired thermal power plants for energy efficiency?

A further review of the seven studies provided insight into how these algorithms were applied in industry to address specific problems by focusing on the datasets, input parameters, training method and results. Furthermore, to provide a more comprehensive overview of the current literature, a detailed overview is provided of each study's contribution and potential future recommendations. A description of each is provided in Table 3.

Table 3 provides an overview of how the different machine learning algorithms were applied in each study, followed by an overview of each study's contribution and future recommendations. The seven studies developed and implemented machine learning models in various forms. As discussed in Section 2, a specific process is followed in training supervised, unsupervised and reinforcement learning algorithms.

Table 3 shows that four studies developed unsupervised learning algorithms, including artificial neural networks and the K-means clustering algorithm. One study applied one supervised algorithm (the least square vector machine), and two focused on deep learning or deep reinforcement learning.

There are similarities between the studies that applied unsupervised learning algorithms. Muhammad Ashraf et al. (2020), Baressi Šegota et al. (2020) and Rossi and Velázquez (2015) used the backpropagation algorithm to train the artificial neural networks. Typical performance metrics used to determine the performance of the ANN included the following: The coefficient of determination, the root mean square error (RMSE), and the mean absolute error (MAE). Most other algorithms followed a similar approach, as explained in Section 2.

Multiple researchers emphasised the value of artificial neural networks. Artificial neural network models are flexible and can be applied to problems even if much of the information is unknown (Smrekar et al. 2009). These algorithms are designed to solve complex problems related explicitly to modelling, control and optimisation (Babatunde 2020). Furthermore, these algorithms can model complex non-linear relationships between inputs and outputs, which is much more challenging than traditional models (Chiroma and Abdulkareem 2016).

Wu et al. (2022) highlighted that the articles identified in their review demonstrated the importance and the applicability of various machine learning algorithms, including decision trees, support vector machines (SVM), and artificial neural networks (ANN), specifically feedforward deep neural networks and recurrent neural networks. Another systematic review conducted by Dhanalakshmi and Ayyanathan (2022) revealed that the artificial neural network (ANN) was mostly applied in the identified studies.

The type of machine learning model employed depends on the volume of data and the problem identified. However, as these different studies are evaluated, it is evident that there is a common thread across all these research studies identified from conducting the review. Specifically, most researchers suggested that artificial neural networks be used when working with large datasets and complex problems, especially those related to industrial processes and power plants.

Evidently, each study contributed to academia and industry by incorporating realistic and practical approaches. The studies by Sleiti, Kapat, and Vesely (2022) and Wahyuningtyas and Asrol (2022) indicated the importance of a decision support system for optimal decision-making. These studies illustrated that a data-driven approach alone is insufficient, and a tool is required to integrate it with system parameters for optimal decision-making. Furthermore, the importance of actual data is highlighted in Table 3 by Sleiti, Kapat, and Vesely (2022) and Wahyuningtyas and Asrol (2022). Real data will improve decision-making and provide an accurate view of the system and its components, specifically when applying machine learning.

5. Summary and conclusions

This research paper investigated machine learning algorithms utilised for energy efficiency improvement at coal-fired thermal power plants by conducting a systematic literature review. Subsequently, this paper provided an overview of the identified relevant studies, focusing on the type of machine learning algorithms and the application thereof. This paper contributed towards research by providing guidelines for sustainable energy supply and improved decision-making towards more efficient processes. Efficient processes result in the reduction of fuel usage, which results in lower emission levels for equivalent power generation capacity. This research paper adds value to the industry and academia by identifying all scientific articles that applied machine learning algorithms in the power industry.

First, an overview of machine learning was provided, explicitly focused on supervised, unsupervised, semi-supervised and deep reinforcement learning algorithms. Second, a systematic literature review followed the same method as Kitchenham et al. (2009). A search protocol was developed by formulating research questions, identifying keywords, databases, inclusion and exclusion criteria, and a quality assessment. From the search process, only 14 studies contained the three keywords. Two studies were identified in Science Direct, 11 in Scopus and one in IEEE Xplore. Seven of these studies were published in 2022, four in 2020, two in 2016 and one in 2015.

From the review, three studies mentioned machine learning: Zima et al. (2022), Variny and Kšiňanová (2022), and Nikula, Ruusunen, and Leiviskä (2016), four conducted a review: Yu et al. (2022), Wu et al. (2022), Dhanalakshmi and Ayyanathan (2022), and Babatunde (2020), and seven applied machine learning at a coal-fired thermal power plant for energy efficiency (Baressi Šegota et al. 2020; Chiroma and Abdulkareem 2016; Fu et al. 2020; Rossi and Velázquez 2015; Sleiti, Kapat, and Vesely 2022; Wahyuningtyas and Asrol 2022). The seven studies applied different machine learning algorithms, including K-means clustering, multiple linear regression (MLR), the least-square vector machine (LSSVM), deep reinforcement learning (RL) and artificial neural networks (ANN). Typical performance metrics used to determine the performance of the ANN included the following: The coefficient of determination, the root mean square error (RMSE), and the mean absolute error (MAE).

Various researchers emphasised the value of artificial neural networks. Artificial neural network models are flexible and can be applied to problems even if all the information is unknown (Smrekar et al. 2009). The models can solve complex problems related explicitly to process modelling, control and optimisation. Furthermore, these algorithms can model complex non-linear relationships between inputs and outputs, which is much more challenging than traditional models (Chiroma and Abdulkareem 2016). The studies by Sleiti, Kapat, and Vesely (2022) and Wahyuningtyas and Asrol (2022) indicated the importance of a decision support system for optimal decision-making. These studies illustrate that a data-driven approach alone is insufficient, and a tool is required to integrate with system parameters for optimal decision-making. The results obtained from this study underline the significance of some applications of machine learning algorithms. For instance, the artificial neural network is famous for detecting non-linear patterns and predicting the behaviour of most industrial processes and power plants (Rossi and Velázquez 2015). The potential future recommendations are identified and explained in Section 6.

6. Limitations and future recommendations

One limitation identified from the systematic literature review conducted in this study is that the keywords are particular, resulting in fewer studies identified on the selected databases. For future research, it is recommended to broaden the search by removing the keyword: coal-fired thermal power plant. By eliminating this keyword, over 3000 studies are identified from Scopus. This will provide insight into how various machine learning algorithms are applied to solve related energy problems in other industries. For example, technologies may be investigated to improve asset failure prediction, resulting in optimised plant production and enhanced energy efficiency.

It is further recommended that a review is conducted that does not only focus on coal-fired thermal power plants but also on the application of machine learning for gas turbines and other fossil fuel plants, including gas-fired and oil-fired power plants. An extensive amount of research has been conducted on machine learning, and a comprehensive review that includes all power plants may further pave the way for future researchers and the industry.

The review showed that various supervised, unsupervised and reinforcement learning models are applied to solve industry-related problems. Based on the systematic literature review conducted and reported on in this paper, limited industry-related applications were identified for machine learning applications at coal-fired thermal power plants under investigation, especially to improve the operational efficiency of the status quo operations. Furthermore, Decision trees, Gaussian processes (GP), and the Naïve Bayes techniques were typically not found in the literature for the research question. It is, therefore, proposed that researchers investigate these techniques for the identified research question, i.e. the aim to improve energy efficiency at coal-fired thermal power plants while reducing fuel usage. Further research should also include comparing these identified techniques with that of an artificial neural network for energy-improved operations at coal-fired thermal power plants. It must be noted that coal-fired refers to coal being utilised within a boiler house to generate steam; therefore, emphasise must be placed on the coal-fired boiler.

Another limitation identified in the literature is applying the long short-term memory network (LSTM) technique. The LSTM has been developed and applied in various industries because it can process time series data (Pan et al. 2021). Furthermore, the LSTM can remember previous information and use this memory to calculate the current output (Pan et al. 2021).

For future research, it is recommended that this modelling technique is applied for optimal operational control of various units at coal-fired thermal power plants. Applying the LSTM may also contribute towards improving the energy efficiency of coal-fired boilers.

Furthermore, neural network predictive control (NNPC) is a model integrated with predictive control (MPC). One of the primary contributors of this model is its ability to be employed in real-time for simple controls and disturbance reduction (Megahed, Abdelkader, and Zakaria 2019). No such model has been identified in this review. For this reason, it may be valuable to investigate the application of such a model to enhance a power plant's automation and operational decision-making. It is recommended that the readers refer to Table 3 for a more comprehensive overview of each study’s potential future research opportunities. These opportunities highlight the limitations of current literature and may be investigated for future research.

Disclosure statement

No potential conflict of interest was reported by the author(s).