Advances in the Application of Machine Learning Techniques for Power System Analytics: A Survey ^†

Abstract

The recent advances in computing technologies and the increasing availability of large amounts of data in smart grids and smart cities are generating new research opportunities in the application of Machine Learning (ML) for improving the observability and efficiency of modern power grids. However, as the number and diversity of ML techniques increase, questions arise about their performance and applicability, and on the most suitable ML method depending on the specific application. Trying to answer these questions, this manuscript presents a systematic review of the state-of-the-art studies implementing ML techniques in the context of power systems, with a specific focus on the analysis of power flows, power quality, photovoltaic systems, intelligent transportation, and load forecasting. The survey investigates, for each of the selected topics, the most recent and promising ML techniques proposed by the literature, by highlighting their main characteristics and relevant results. The review revealed that, when compared to traditional approaches, ML algorithms can handle massive quantities of data with high dimensionality, by allowing the identification of hidden characteristics of (even) complex systems. In particular, even though very different techniques can be used for each application, hybrid models generally show better performances when compared to single ML-based models.

1. Introduction

The power system management and development have constantly been changing due to expanding complexity and distributed modern power networks. [1]. Principally, the increasing distribution of Renewable Energy Sources (RESs) with intermittent energy generation and technological novelties in power system management and control demand reliable power predictions and more precise monitoring models [2,3]. In recent years, researchers developed advanced solutions based on Machine Learning (ML) algorithms to solve the bottleneck of conventional lumped parameter simulations. In practice, conventional traditional simulation techniques based on deterministic methods are still dominated in power grids. However, the high performance of machine learning solutions in terms of accuracy computational speed, and scalability brings novelties in power grids management and control. Therefore, it is expected to boost the adaptation of these techniques for short- to medium-term forecasts of the power grid system operation to meet this gap while getting benefits of advantages of traditional approaches.

As the high-quality sensor prices–such as smart meters, Phasor Measurement Units (PMUs), and Remote Terminal Units (RTUs), or other measurement devices—are constantly decreased, they are increasingly distributed in the power system and are continuously acquiring a massive amount of heterogeneous datasets. Analyzing and processing all these data provides new insights and advances in the control and operation of smart grids thanks to innovation in ML and big data frameworks to handle structured and unstructured data. Traditional time-domain methods are computationally inefficient; thus, they are not good candidates for real-time applications in which response time is a decisive concern [4,5]. The expected significant penetration of Electric Vehicles (EV) charging stations, and the increasing expansion of the Internet of Things (IoT) devices in private and public sectors such as smart buildings introduce new challenges and opportunities for the perception of accurate Day-Ahead Load Forecasts (DALFs) in micro-and smart- grids [6,7]. At the same time, the transition towards decarbonization of systems leads to the integration of distributed energy systems which usually generate energy in an intermittent and stochastic manner, such as wind or solar energy generators with no inertia. Consequently, the growing complexity of power flow patterns requires novel approaches to render reliable, efficient, and economical solutions.

In this context, advanced machine learning models have been shown promising results to provide new valuable knowledge and insights and identify hidden data patterns, trends, and relationships. In [8], the authors briefly summarized the ML paradigm and presented the literature review on applications of ML methods in Power systems till the end of 2017. This paper continues the authors’ work presented at [8] and aims at providing a systematic review of the various machine learning algorithms used to analyze, monitor, and model power flows, power quality events, photovoltaic systems, intelligent transportation systems, and load forecasting services. The authors selected the Journal papers for literature review based on publication date, number of citations, and novelty in contributions. The main contribution of this article is as follows:

The ML paradigm and well-known ML algorithms are categorized and presented;

The systematic review summarizes not only the main contributions of each article but also provides information regarding the explicit application, data source, and models, by mainly considering articles published since 2018; this study used Google Scholar, Scopus, IEEE Xplore, and the MPDI databases for literature review, which ended in February 2021.

To make a fair comparison between models, the characteristics of a standard dataset for the testing of the reviewed models are presented.

The remainder of the article is structured as follows: Section 2 explains the machine learning paradigms, well-known algorithms, and performance metrics. A literature survey on recent advanced machine learning applications in power flow, power quality, photovoltaic systems, electric transportation systems, and load forecasting is presented in Section 3. Section 4 discusses the results and notes achieved in the literature review, and Section 5 summarizes the final remarks and conclusions.

2. Machine Learning

Artificial Intelligence (AI) deals with the broad topic related to the perception and extraction of knowledge from data. AI can be divided into two main subsets: machine learning and deep learning. Machine learning is the main subset of artificial intelligence, while deep learning can be represented as a subset of machine learning.

Machine learning is an interdisciplinary research field that consolidates expertise and knowledge from diverse areas and aims at proposing solutions to given problems that can be used to reply to similar questions raised by different contexts. More specifically, machine learning is the subset of AI that deals with the extraction of knowledge from the experience by analyzing and manipulating data gathered from real-world use cases. The primary purpose of machine learning is to develop reliable active learning models equipped with computerized patterns learning from raw data and perform fast-response predictions applied in decision-making processes [9,10].

Deep learning and neural networks are the most famous machine learning subset. Thanks to the use of (typically) multi-layered Artificial Neural Networks (ANNs), deep learning can handle unstructured datasets and can recognize complex input data patterns. In deep learning, different architectures can be designed using neural unit cells in various layers, unless other machine learning algorithms are fixed.

Figure 1 illustrates artificial intelligence, machine learning, and deep learning concepts in the schematic description by means of a Venn diagram.

2.1. Machine Learning Paradigms

In machine learning, training a model intends to learn the values of the parameters (or weights) and the bias from input data, while in traditional methods (i.e., with predefined algorithms), both the model and its parameters are given to a computer to perform a task. Labeled data are samples with a sort of meaningful “tag”, “label”, or “class” that are informative or desirable to know—for example, whether an Alternating Current (AC) power signal contains harmonic distortion(s) or not. In contrast, unlabeled data are samples with no explanation; in other words, it has only row data without any “tag” or “label” assigned to it—for instance, voltage and current signals of an electric motor.

Machine learning tasks are principally arranged into three main classes: supervised, unsupervised, and semi-supervised learning. Supervised learning algorithms work with labeled data with the objective of mapping new input data to the known target output values. On the contrary, unsupervised learning models process an unlabeled dataset, in which target values are unknown, to draw insights by learning hidden complicated patterns and structures spontaneously. Semi-supervised algorithms deal with a dataset that some samples are labeled, and more extensive samples are unlabeled. These algorithms are designed to benefit from both advantages of supervised and unsupervised methods [11];

Supervised learning is categorized into classification and regression problems. A classification problem predicts output variables as a category, such as “cat” or “dog.” Contrarily, in regression problems output variables are numerical values [12];

Unsupervised learning algorithms are generally divided into clustering or dimensionality reduction (or sometimes called embedding) methods [11]. For instance, in anomaly detection, a clustering algorithm is applied to data to identify false data by scanning outliers in a dataset or noticing abnormal patterns;

Semi-supervised learning makes use of the mixture of labeled and unlabeled data as the training dataset. Semi-supervised models act as active learners [13]. There are two main semi-supervised learning algorithms, namely reinforcement learning and Generative Adversarial Networks (GANs). In reinforcement methods, if a model does a task correctly, it would get a reward. The objective of reinforcement learners is to build a model to maximize rewards through an iterative process [14]. Reinforcement learning is suitable for an interactive or dynamic environment that a model can improve itself based on policies defined by an expert, for example, playing a game or self-driving cars. GANs generate models based on deep neural learning methods to discover and learn patterns of input data. Then, the generative model can be used to create new data examples that resemble a training dataset. For instance, GANs can create pictures that look like human faces images, even though the faces don’t relate to any actual person.

2.2. Machine Learning Algorithms

Many different machine learning techniques have been proposed in recent years, particularly consisting of hybrid ML-based models, making use of two or more machine learning techniques or even other statistical or mathematical models. For example, ensemble learning models include different weak learners such as decision trees, support vector machines, and linear or logistic regression. This section discusses the basic and most relevant machine learning techniques in each category.

2.2.1. Classification Algorithms

There are several classification algorithms; the most commonly used ones are presented as follows [15]:

Logistic Regression (LR): LR is widely used for binary classification tasks where an output belongs to one class or another (0 or 1). In this algorithm, a threshold is defined to indicate examples will be labeled into which class using hypothesis and logistic function (usually sigmoid curve). The hypothesis determines the likelihood of events to generate data and fit them into the logarithm function that forms an S-shaped curve called sigmoid. Then, the logarithm function is used to predict the class of new inputs. Even though logistic regression provides better performance in binary classification tasks, it can also be used in multiclass classification problems, by applying the one versus all strategy [16];

K-Nearest Neighbors (KNN): this algorithm is one of the most basic yet broadly used classifiers. It is generally used to find data with similar characteristics and group them in the same class, without making any assumptions on data distribution. The groups are constructed by considering the attributes of the neighboring samples. It is used in real-life problems in several applications such as data mining, pattern recognition, and invasion detection [17,18];

Naïve Bayes (NB): this technique is one of the most powerful classification algorithms based on an extension of Bayes’ theorem, assuming each feature is independent to capture input-output relationships. Bayes’ theorem compares the probability of an event happening to what has already happened, for example, the probability of having a fire (event A) while the weather is hot (event B, which is present) [19]. The naïve algorithm is simple to implement and can easily predict labels of new inputs. Additionally, when domain knowledge confirms the feature independence, with less data, it has a better performance than other classification algorithms such as logistic regression. On the other hand, in real life, it is not easy to have data with entirely independent features; moreover, when there is an input that was not followed up in the training phase, the algorithm assigns zero probability, and it does not classify this input in any group. This technique is used in various applications such as text classification and spam filtering [20];

Support Vector Machine (SVM): This algorithm is widely used in classification tasks and also applied in regression problems. The main idea of SVM is to transfer data to higher n-dimensional space to find an ideal hyperplane to differentiate classes [21]. In simple words, these support vectors are coordinates of a new n-dimensional coordinate system. This method is commonly used in binary classification, but it is computationally expensive and slow in the big data domain;

Decision Tree (DT): This algorithm is based on different hierarchical steps that lead to certain decisions. It applies a treelike structure to represent decision paths with induction and pruning steps. In the induction step, the tree structure is built, while, in the pruning step, the complexities of the tree are reduced. The inputs are mapped to outputs by traversing each path through different branches of the tree [22]. DT is a powerful classification tool, simple to structure and with good performance. However, with even small variations in data, DT can become unstable. Furthermore, it can easily become overfitted, especially in a thorny tree with many branches and conditions, thus, it does not generalize well on new inputs. Regularization, bagging, and boosting techniques are usually used to avoid overfitting problems in the DT [23];

Random Forest (RF): This classifier is very similar to the decision tree. Compare to DT, RF uses several decision trees, instead of having only one tree. This technique can be applied in massive data set to classify data or measure the importance of each feature in the final decision. In many applications, the random forest is preferred over the decision tree because it can be more accurate and overcomes the overfitting issued of DT. However, this technique is not easy to implement since it has a complex structure, and it is not recommended for real-time prediction purposes because it is generally slower than other models [24].

2.2.2. Regression Algorithms

Several regression algorithms (numerical or continuous value prediction) have been introduced in the scientific literature; the most commonly used ones are presented in the following:

Linear Regression (LR): this technique tries to find the fittest straight hyperplane to the data [25]. It is commonly used when there are linear relationships between variables, and it can avoid overfitting by regularization techniques such as LASSO, Elastic-Net, and Ridge [26]. However, it is not flexible in finding the best solution for non-linear relationships in variables and complex patterns;

Regression Tree (RT): This technique has the same hierarchical structure as the decision tree, but it takes numerical values as input. The branching procedure not only maximizes the learning gain but also learns non-linear relationships between variables. Even if this method is robust to outliers and easy to implement, it is prone to overfitting problems [27]. In addition to the regression tree, random forests and Gradient Boosted Trees (GBM), which are the most commonly used ensemble methods, are also applied in numerical predictions and have better performance concerning overfitting issues;

Deep Neural Network (DNN): Deep neural network, or multi-layer neural network, is widely used in several domains. Indeed, thanks to their ability to capture complex patterns, DNNs can be used both as regression algorithms and classifiers. The non-linear relationships between features are learned by non-linear activation functions and hidden layers between the input and the output [28]. There are several techniques and methods to improve the performance of neural networks, as well as different advanced neural network-based models such as Convolutional Neural Network (CNN) or Recurrent Neural Network (RNN) [29,30]. Different from other algorithms, in DNNs, a deep knowledge of how to tune the parameters of the neural network is required to develop a working neural network model. In addition, even though neural network models work well in the big data domain, they are usually very computationally expensive methods;

Extreme Learning Machine (ELM) has a wide range of applications in the data-driven approach. ELM has been used in regression, classification, clustering, sparse representation, feature extraction or learning, and compression. This feedforward neural network does not apply the backpropagation gradient-based mechanism to update the network weighted values; instead, it randomly assigns random values to the weight and bias terms of the network [31]. The main advantages of this kind of algorithm are (i) the faster training phase and (ii) the better interpolation results. On the contrary, the accuracy results of ELM is not promising, even if compared to basic MLP models;

Support Vector Machine or kernel SVM can also be used for regression problems, even though it is mostly used in classification problems;

XGboost, finally, is a (recently) widely used rugged decision-tree- and ensemble-based algorithm with a framework that is designed considering a gradient boosting procedure [32].

2.2.3. Clustering Algorithms

Clustering techniques try to group instances with the same properties in the same cluster. These techniques are commonly used in other fields than machine learning, such as image analysis, pattern recognition, data compression, and statistical analysis. The most well-known algorithms are as follows:

K-means: this technique, one of the simplest and intuitive machine learning algorithms, separates instances in the k centroids or clusters with equal variance. After selecting the number of clusters (K), the algorithm finds the best k clusters by minimizing the criterion known as inertia through the iterative procedure and changing the position of centroids [33]. As it is simple to interpolate and scales well to big data, it has been applied across a wide range of applications in various domains;

DBSCAN: Density-Based Spatial Clustering of Applications with Noise (DBSCAN), which is widely used in data mining and machine learning, finds core instances of high density and extends clusters with the specified radius (usually Euclidean distance) around them. Low-density regions are distinguished as outliers. The primary problem in DBSCAN is selecting clustering attributes, detecting noise with varied densities, and significant differences of amounts of boundary objects in opposed directions of the corresponding clusters [34]. The smallest number of instances to constitute a dense region and how close instances should be to each other in the same region are defined by an expert. Even though this algorithm, which is a very popular clustering technique, is widely used, it badly behaves with very sparse or high dimensional datasets;

Spectral: this clustering algorithm, which is also an exploratory data analysis technique, performs dimensionality reduction through eigenvalues (spectrum) of the similarity of data instances, by then grouping similar data instances with reduced dimensions into the same cluster [35]. This approach is practically applied when the center of clusters and their spread does not appropriately describe the whole cluster (non-convex cluster), such as in image segmentation problems. The spectral technique is widely used because it is a fast response technique and outperforms other clustering techniques, especially in sparse datasets.

2.2.4. Embedding Algorithms

In many cases, especially in the big data domain, the presence of a large number of variables or features in a dataset makes it difficult to interpret the relationship between them. Training a model on the whole dataset could easily make the model not sufficiently generalized on new unseen data (overfitting problem). Embedded Algorithms (EAs) can be applied to extract new features from data without losing essential information before implementing sophisticated ML models. EA techniques could also be used directly for prediction purposes. Embedding algorithms can be subdivided as follows:

Principal Component Analysis (PCA): the main aim of PCA is to reduce high-dimensional datasets to a smaller dimension. PCA projects each data instance onto the main components or ranks while retaining as much data variation as possible. PCA techniques, such as Singular Value Decomposition (SVD), use eigenvectors of the covariance matrix of data to reduce the dimension of the dataset or making a prediction [36];

Autoencoder: this is one of the current states of the art techniques leveraging neural networks. Autoencoders are widely used in different applications, such as data compression. The autoencoder learns a representation (encoding) of the input dataset and ignores noise through embedding architecture, and reconstructs the input data as close as possible to its actual forms (decoder). A typical autoencoder consists of three parts, namely: an encoder, a bottleneck, and a decoder [37]. The encoder tries to compress the data to a lower dimension with the best representative, the decoder attempts to regenerate an input by eliminating the noise in the dataset, while the embedded data is stored in the bottleneck. It is possible to use the encoder part of a well-trained autoencoder for dimensionality reduction, or use the whole model, for example, in anomaly detection [38].

Figure 2 summarizes the different machine learning paradigms and techniques used in power system analytics, by providing examples for each category.

2.3. Model Performance Evaluation Metrics

The metrics that are used in each machine learning algorithm are different from each other. In

Table 1. Model Performance Evaluation Metrics.

Discrete		Continuous
Metric	Formula	Metric	Formula
Accuracy	$\frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}$	Mean Square Error (MSE)	$\frac{1}{n} {\displaystyle \sum }{\left(y-\widehat{y}\right)}^2$
Error	$\frac{\mathrm{F}\mathrm{N}+\mathrm{F}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}$	Root Mean Squared Error (RMSE)	$\sqrt{\frac{1}{n} {\displaystyle \sum }{\left(y-\widehat{y}\right)}^2}$
Precision	$\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}$	Mean Absolute Error (MAE)	$\frac{1}{n} {\displaystyle \sum }\left|y-\widehat{y}\right|$
Recall	$\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$	Mean Absolute Percentage Error (MAPE)	$\frac{1}{n} {\displaystyle \sum }\left|\frac{y-\widehat{y}}{y}\right|$
F1	$\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$	R-squared (R²)	$1-\frac{{\displaystyle \sum }{\left(y-\widehat{y}\right)}^2}{{\displaystyle \sum }{\left(y-\overline{y}\right)}^2}$

the most used metrics in discrete and continuous cases are discussed. In this table, True Positive (TP) and True Negative (TN) are samples that are correctly predicted as positive and negative, respectively. In contrast, False Positive (FP) and False Negative (FN) are samples that are incorrectly predicted as positive and negative, respectively. In continuous metrics, $y$ is the actual value, $\widehat{y}$ is the forecasted amount, and $n$ is the number of prediction samples.

3. Literature Review

Machine learning is widely applied to address various problems to bring novel solutions or improve the performance of existing applications. The main state-of-the-art machine learning-based applications in power systems are in power flow, power quality, photovoltaic system, intelligent transportation, and load forecasting.

3.1. Power Flow Applications

Compared to traditional algorithms, machine learning technologies make power flow problems easier to be handled. For example, algorithms like CNN, KNN, SVM, reinforcement learning, and decision tree affected power flow optimization problems in terms of accuracy, computational speed, and response time.

Table 2. Overview of Research using Machine Learning for Power Flow.

Reference	Year	Application	Data	Method(s)	Remark(s)/Contribution(s)
Lei et al. [39]	2020	Optimal Power Flow (OPF)	IEEE 39, 57, 118-bus, and Polish 2383-bus (wind and photovoltaic power connected) 30,000 samples for training, and 10,000 samples for validation	ELM	Decompose OPF via a data-driven regression framework with three stages stacked extreme learning machine (SELM); implemented the multiple supervised layers with reinforcement mode with an overall 98.71% accuracy rate, significantly higher than benchmarks.
Wang et al. [40]	2020	Online Detection of Geomagnetically Induced Currents in Power Grids	Simulation data based on real-life power grid operation (10,800 samples for training, 1200 samples for validation, and 6000 samples for testing)	CNN	Developed hybrid feature extraction consists of pseudo-continuous wavelet transform (PCWT) and short-time Fourier transform (STFT); improved overall detection accuracy to 90.15% for different noise levels and achieved the detection results within 30 m
Ravikumar et al. [41]	2020	Anomaly Detection and Mitigation (ADM) for Wide-Area Damping Control	Synchrophasor measurement data and simulated the cyber-physical system (CPS) dataset (60 fps transmission rate)	KNN and PCA	Improved efficiency of ADM with domain-specific features extraction and selection via Teager–Kaiser Energy Operator (TKEO), Principal Component Analysis (PCA), Wide-Area System Measures (WASMs), and Primitive Measures; Proposed KNN-based model with a 95.5% accuracy rate, better than other ML-based models.
Zhang et al. [42]	2021	Volt-VAR Optimization in Smart Distribution Systems	Unbalanced IEEE 13-bus and 123-bus systems (9000 operating conditions for training, and 13,000 operating conditions for testing)	Reinforcement learning	The improved accuracy rate of voltage regulation with an average of 99.80% compared to 90.02% of baselines; Achieved an average executive time of 21.7 and 46.2 ms for 13-bus and 123-bus systems, respectively, and 28.38% for the loss reduction percentage.
Baker et al. [43]	2019	Joint Chance Constraints in AC Optimal Power Flow	IEEE 37-node test feeder (5-min data from August 2012 weekdays) 1152 samples for training (4 days), and 864 samples for testing (3 days)	SVM	Improved classic methods based on union bound (or Boole’s inequality) and the accuracy rate with 0.19% and 4.73% error rate for false classification of binding and non-binding events, respectively.
Li et al. [44]	2020	Transient Stability Assessment of Power System	IEEE New England 10-machine 39-bus system, IEEE 16-machine 68-bus system, and IEEE 47-machine 140-bus system (5984, 11,792, and 29,520 samples, respectively) 6-s simulation time per instance, and 0.01-s step	XGBoost and FM	Proposed hybrid XGBoost-FM model robust to noise. Used extreme gradient boosting (XGBoost) for automatic feature builder, and factorization machine (FM) as a classifier with the enhanced detection time of 0.9349 s; improved accuracy by using both original and artificial features to 98.21%.
Hong et al. [45]	2020	State Estimation of Distribution Network	IEEE 13-, 34-, and 37-node test feeders (240 samples for training, and 60 samples for testing)	LR, SVM, and FFNN	Estimated the voltage magnitudes and angles of several successive buses with 0.01 p.u. and 0.189° error respectively; SVM outperformed LR and FFNN, especially when the relationship between inputs and outputs is unknown, the input bus was missing, there is a measurement error, and using few adjacent buses as input buses.
Karagiannopoulos et al. [46]	2019	Optimized Local Control for Active Distribution Grids	Seasonal historical data (30-day dataset; 1-h time resolution; 7200 samples)	SVM	Proposed Data-driven method to obtain local Distributed Energy Resources (DERs) controls without monitoring and communication infrastructure; outperformed standard industry local control with an overall RMSE accuracy of 0.158.
Zhao et al. [47]	2020	Real-Time Power Grid Multi-Line Outage Identification	IEEE 30, 118, and 300 bus systems (300,000, 800,000, and 2.2 million data generated, respectively)	ANN	Generated a large number of samples with Monte Carlo simulation with full-blown power flow models; Achieved an overall classification accuracy of 99%, and outstanding performance in recognizing multi-line outages in real-time with a small amount of data.
King et al. [48]	2015	Algorithm Selection for Power Flow Management	IEEE 14-bus (10,000 states for testing), IEEE 57-bus (10,000 states for testing), and a real 33-kV distribution networks (17,520 states for testing -a year of half-hourly profile data-)	ANN, DT and RF	Shown that ML-based methods can create effective algorithm selectors for power flow management based on algorithms’ behavior data within 1 ms for future complex networks
Labed et al. [49]	2019	Overloaded Power System Alleviation	Algerian (Adrar) 22-bus system (75% of data for training, and the remaining 25% for validation)	ELM	The proposed method outperforms SVM and ANN learning algorithms with 1.9465*10 MSE accuracy and 0.0023 s response time on the testing phase with generalization performance; this fast time response minimized the threat and risk of outage and cascade failure.

elaborates more into detail the recent advancements in machine learning applications in power flow.

3.2. Power Quality Applications

The power quality, one of the most critical topics in electrical systems, has also been affected by machine learning, which can be used to improve speed and accuracy in disturbances detection, or distortions classification, and estimations for future cycles. In addition, ML can also be used on a wide set of PQ parameters related to load functioning such as active power, reactive power, complex power, fundamental frequency, and power factor.

Table 3. Overview of Research Using Machine Learning for Power Quality.

Reference	Year	Application	Data	Method(s)	Remark(s)/Contribution(s)
Ray et al. [50]	2018	PQ Disturbances Classification in Solar PV Integrated Microgrid	Generated dataset from solar PV integrated microgrid model (600 samples; 5-kHz sampling frequency)	SVM, ICA	Proposed the independent component analysis (ICA) and statistical feature extraction using SVM; ICA-SVM improved accuracy to 99.5% compared to 97.8% of traditional Wavelet transform-SVM
Sahani et al. [51]	2020	A Real-time Power Quality Events (PQEs) Recognition	Synthetic (50 samples for training, and 100 samples for validation) and real (100 samples per distortion –validation-) power quality events data	ELM	Robust anti-noise online PQEs classification; Outperform other models with 98.86% accuracy rate and 0.019 s response time
Turovic et al. [52]	2019	PQ Distortions Detection in Distribution Gird	IEEE 13-bus system modified with DG (85% of the samples for training, and 15% for validation)	ANN, SVM, and KNN	Detection speed comparison between ML algorithms and traditional FFT; ANN has the best detection’s speed with 0.432 ms (600% more than FFT) with a 99.41% accuracy rate
Liao et al. [53]	2018	Voltage Sag Estimation	IEEE 68-bus test network (374,400 faults simulated)	CNN	Automatic system area mapping and feature extraction in the input bus matrix from various local areas in the power network; reached 99.41% of overall estimation accuracy.
Vantuch et al. [54]	2017	PQ Forecasting for Off-Grid System	Experimental off-grid laboratory (141,537 one-minute-resolution measurements, more than 3 months) and simulated data	RF	More than overall 90% accuracy for forecasting short-term (15 min ahead) PQ disturbances
Bagheri et al. [55]	2018	Voltage Dip Classification	6000 real measured voltage dips data over different countries One month of recording	CNN	Developed a robust automatic feature extraction using a space phasor model (SPM) and CNN; outperformed the other existing models with a 97.72% accuracy rate and 0.50% false alarm
Sahani et al. [56]	2020	Power Quality Events Recognition	Synthetic and laboratory PQDs (150 samples per class −50 for training and 100 for validation-; 3.2-kHz sampling frequency)	ELM and VDM	Developed an automatic PQEs patterns recognition system from nonstationary PQ data by using integrating variational mode decomposition (VMD) and Online P-Norm Adaptive Extreme Learning Machine (OPAELM); shorter event recognition time and classification accuracy rate of 99.3%.
Wang et al. [57]	2019	Power Quality Disturbance Classification	Synthetic data (16 PQDs) IEEE Power Engineering Society database (1000 samples; The influence of data imbalance is eliminated by applying an enhancement process)	ELM	Select less than 10 features out of 4500*1280 signal matrix via discrete wavelet transform (DWT) feature extraction and particle swarm optimization (PSO) feature selection; Proposed PSO hierarchical ELM (PSO-H-ELM) classification with automatic encoders and sparse constraints; overall classification accuracy rate is above 95%, and high calculation speed (less than 0.169 s).
Shen et al. [58]	2019	Detection and Classification of PQDs in Wind-Grid Distribution Systems	Synthetic data (2400 samples) Simulated data from the standard IEEE 13 node bus system with wind-grid distribution (5590 samples) 10-kHz sampling frequency	CNN and IPCA	Used Improved Principal Component Analysis (IPCA) for extracting statistical features; applied 1D-CNN classification, which gives 99.76% accuracy on average for different noise levels, higher than other classification methods.
Deng et al. [59]	2019	Type Recognition and Time Location of Combined Power Quality Disturbance	Synthetic data from IEEE 1159 power quality standard for training (1000 samples × 96 combinations of PQD) and real data generated in a lab for testing (140 samples)	GRU	Proposed bi-directional GRU model for classifying 96 different kinds of disturbances noiseless and with noise from 10 dB to 50 dB; have a 98% accuracy level on real operational data and the absolute error of starting-ending times location less than 0.469 ms.
Cao et al. [60]	2019	Transient voltage stability analysis based on frequency, active power, and reactive power	Simulated data of different nodes collected by phasor measurement units	CNN and Deep Learning	Decision optimization algorithm based on PQ parameters implemented; reactive power compensation decision based on deep learning performed
Abed [61]	2018	Power factor enhance and control	Simulated power system	Clustering neural network	The proposed method allows improving power factor
Zhang et al. [62]	2020	Reactive load prediction	SCADA data from a real power grid (357 busloads) Training set: data from June 1 to August 5 Test set: data from August 6 to August 22. 15-min sampling period	Deep learning	Reactive power load of buses can be accurately predicted; accuracy is better than that obtained with other prediction models; result of great significance for reactive voltage control
Nakawiro [63]	2020	Voltage and reactive power control	Simulated dataset of grip operation (on-load tap changer, load, and wind power) 1 year of operation (hourly data)	DT and KNN	The highest classification accuracy is achieved with a DT; accuracies obtained in the simulations are satisfactory for some classes; performance heavily relies on the distribution of the target output and number of samples per class
Moreira et al. [64]	2018	Power factor compensator (based on PQ parameters: power factor, unbalance factor, harmonic distortion, reactive power, etc.)	Training: Simulations characterized by a human specialist (1,355,154 samples per disturbing load). Real measurements added Test: IEEE 13-bus (111,055 samples). Three real test sets (disturbing loads)	DT, KNN, SVM, and ANN	PQ parameters used to analyze the functioning of a power system; DT is highly effective in classification; 100% accuracy achieved
Valenti et al. [65]	2018	Non-intrusive load monitoring based on active and reactive power	Two public datasets: Twenty-one power meters; 60-s sampling period; 2 years monitoring. Four different locations; multiple sampling frequencies	ANN	Introducing reactive power increases F1 score performance from +4.9% to +8.4%; reactive power provides significant information for non-intrusive load monitoring

summarizes the most recent improvements and achievements in the use of ML techniques in power quality applications.

3.3. Photovoltaic System Applications

Machine learning algorithms have been widely used for different purposes in Photovoltaic (PV), from forecasting the long-, medium-, and short-term energy generation, to fault detection and classification. The most recent works in this field are summarized in

Table 4. Overview of Research Using Machine Learning for Photovoltaics Systems.

Reference	Year	Application	Data	Method(s)	Remark(s)/Contribution(s)
Keerthisinghe et al. [66]	2020	PV Forecasts for Capacity Firming	Dataset of 2013–2018 coming from empirical formula and 2019–2020 real-data of Arlington Microgrid; input frequency every half an hour for one day and two samples for output	LSTM	Proposed encoder-decoder LSTM-based model for short-term (1-h ahead) PV generation prediction resulted in reducing the yearly battery energy throughput by 29% and the number of battery cycles with a greater than 10% depth-of-discharge by 51%.
Wen et al. [67]	2019	PV Prediction in Ship Onboard	Historical hourly data of meteorological information along with the ship route movement for a year	ELM	The proposed ML-based model with the particle swarm optimization (PSO) has a MAPE accuracy level of 25.41% in the training phase for five-hour ahead prediction; the difference between prediction and experimental results has 14.96% of the absolute error in the test phase, which means it has a high potential in practical cases.
Dhibi et al. [68]	2020	Fault Detection and Classification in Grid-Tied PV System	Emulate the operational real PV array dataset using Chroma 62150H-1000S programmable dc power supply; 100 $\mathsf{\mu }\mathrm{s}$ sampling time with 1501 samples for 6 different classes for both training and testing	RF and K-means	Proposed two classifiers based on Reduced kernel RF for detecting faults: Euclidean distance-based RK-RF and K-means clustering-based RK-RF with 100% accuracy and reduced computational time 65.16% and 53.33 compared to kernel RF, respectively; redundancy between samples was reduced by using Euclidean distance as a dissimilarity metric; the K-means clustering method used to reduce the training data amount.
Zhang et al. [69]	2020	Day-Ahead PV Estimation	Real datasets from Cupertino, CA, USA, from July 2015 to December 2016; and Catania, Sicily, Italy, from January 2011 to December 2011 with a 15-min sampling rate	LSTM and AE	Improved the prediction accuracy to 8.39% nRMSE compared to benchmarks with the proposed hybrid Auto Encoder (AE) LSTM model for three months testing; the proposed persistence model (PM) has a high accuracy of 0.72% nRMSE for consecutive clear days; Applied the Root Mean Squared Euclidean Distance Difference (RMSEDD) to extract and select the most valuable features to increase the model accuracy.
Chang et al. [70]	2020	Short-term Photovoltaic Power Prediction for Edge Computing	Real PV output and PV meteorological dataset; one sample every 30 min	LightGBM	Proposed a tree-structured self-organizing map (TS-SOM) algorithm for clustering weather; used Bayesian optimization algorithm is employed for temporal pattern aggregation to determine the optimal size of time steps; the proposed LightGBM outperforms other algorithms in training and execution time (0.020 s) with 35.49 RMSE accuracy, suited for edge computing devices.
Khan et al. [71]	2020	Islanding Classification for Grid-Connected PV	Simulation data; total size equals 4526 samples and 7 features; 3168 samples for training and 679 samples for testing	ANN	Proposed islanding detection model-based Wavelet transform for feature extraction and Multi-layered Perceptron (MLP) for classification with 97.8% accuracy under 0.2 s on unseen conditions.
Wang et al. [72]	2019	Key Weather Factors from Analytical Modeling Toward Improved PV Forecasting	Real hourly dataset for a year of three PV arrays in Australia from April 2012 to June 2013 with 11 independent variables	SVM, ANN, and KNN	Improved the accuracy level for each season by using PCA for feature extraction and KNN for classifying the prediction period into the historical periods with the most similar weather situations; for example, on sunny days, with the proposed method, SVM has 3.97 instead of 8.14, ANN has 4.09 instead of 8.45, and weighted KNN has 8.86 instead of 9.33 nRMSE accuracy; this method helps ANN converges much faster with 37.72% computational time reduction.
Gao et al. [73]	2020	Fault Identification Method for Photovoltaic Array	Simulation dataset with 1320 samples and experimental dataset with 1892 samples with a ratio of 6:2:2 for training, validation, and testing	CNN and GRU	Outperformed benchmark methods with 98.41% accuracy in 28.1 ms detection time using CNN as automatic feature extractions and Residual-GRU for memorizing time-series dynamic features; outperformed benchmarks also in the presence of 10 dB to 50 dB noise level; reached accuracy of 95.23% when some features are missing (temperatures and irradiances).
Catalina et al. [74]	2020	PV Energy Nowcasting	Hourly satellite and Numerical Weather Predictions (NWP) dataset with 4645 sample size for 2015	SVR	Proposed Gaussian SVRs models using satellite data and NWP information to improve the PV energy nowcasting in the three real experimental regions.
Ray et al. [75]	2020	Long-term PV Output Forecasting	Historical hourly datasets of 24 years of four different locations in North Queensland in Australia; dataset from 1990 to 2013 was used for training and 2014 for testing	LSTM and CNN	The proposed hybrid model, consisting of CNN and LSTM, outperforms other methods with RMSE lower than 15 for all studied locations and low computational cost (203.63 s) for training and prediction.
Yap et al. [76]	2020	Grid Integration of PV	Simulation dataset with 0.1 s sampling time	Reinforcement learning	Proposed the new virtual inertia control algorithm for integrating PV to a grid with higher frequency nadir, lower frequency deviation (reduced by 0.1 Hz), smaller steady-state error (reduced by 27%), faster settling time (reduced by 35%), lesser active power injection or absorption, and lesser overshooting compared to traditional approaches.
Keerthisinghe et al. [77]	2019	Energy Management of PV-Storage Systems	Historical and one year-long simulation datasets with 30 min time intervals for each day	ANN	Proposed an ANN-based model based on dynamic programming (DP), which, compared to other methods, has better quality and faster response time (27.15 s); this method reduced a daily and yearly electric cost by more than 50% for four different scenarios considering PV output, electrical demand, electricity price, and battery SOC.

.

3.4. Intelligent Transportation Applications

Artificial intelligence, especially machine learning applications, are widely used in intelligent transportation, to develop smart online traffic management systems, from safety applications (e.g., driving distraction detection) to optimized traffic scheduling. Self-driving cars, for instance, have been recently developed only thanks to the advancements in machine learning.

Table 5. Overview of Research using Machine Learning for Transportation.

Reference	Year	Application	Data	Method(s)	Remark(s)/Contribution(s)
Ashqar et al. [78]	2019	Transportation Mode Recognition	Real data of GPS, accelerometer, gyroscope, and rotation vector sensors through a smartphone app for 10 travelers with 25 Hz sampling frequency	Two-layer hierarchical framework RF-SVM	Introduced new extracted frequency domain features and increased accuracy rate to 97.02% compared to 95.10% of traditional approaches
Jia [79]	2019	Analysis of Alternative Fuel Vehicle Adoption	Person-, household-, trip- and vehicle real-dataset (2017 NHTS Dataset) from April 2016 to April 2017	RF	Extraction influencing factors from large-scale 2017 NHTS Dataset and Categorized them; RF outperformed other models (LR, NB, SVM, and DT) with good accuracy (97.99%) and high AUC value for adoption prediction.
Aksjonov et al. [80]	2019	Detection and Evaluation of Driver Distraction	Simulation data: speed limit, a radius of the road, lane-keeping offset, and vehicle speed for 18 subjects with 50 Hz sampling frequency	Nonlinear regression based on Euclidean distance and Fuzzy logic	The proposed method improved the RMSE accuracy level from 2.1345 to 1.9992 for speed and 0.1506 to 0.1405 for distance. Training time also decreased from 148.072 to 96.150 s compared to the standard ANFIS predictor
Nallaperuma et al. [81]	2019	Online Smart Traffic Management	Real-time Bluetooth sensor network data and social media data (Twitter) from the arterial road network in Victoria, Australia; 24 and 7 days data for training and testing, respectively, with data horizon equal to 15 min	LSTM and Reinforcement learning	Short-term traffic flow with normal fluctuation prediction with 0.0727 MAE accuracy; overcome the limitation of labeling data and strict assumptions regarding data and traffic behaviors.
Gjoreski et al. [82]	2020	Monitoring Driver Distractions	Real data of 68 people through physiological sensors, the emotional response, and facial-expression extraction with 1 Hz sampling frequency	Comparison of classical ML and deep learning algorithms	The classical extreme gradient boosting (XGB) outperforms the deep learning method with 94% F1-score accuracy compared to 87% for classifying complete driving sessions.
Li et al. [83]	2019	Security: SQL Injection Detection	Real-data and data augmentation from enterprises and various social platforms; 36,422 real samples and 30,000 generated samples	Deep LSTM network	Overcome the overfitting problem and increase accuracy (93.47–99.58%) due to data augmentation compared to the shallow and deep ML algorithms.
Ou et al. [84]	2019	Real-Time Estimation of Dynamic Origin-Destination flow	Generate training dataset from real-traffic dataset and traffic survey, and testing on real-time data with 15-min intervals sampling for 15 days in June 2017	CNN	Capture the dynamic mapping patterns and reconstruct trajectories with MAPE average accuracy less than 5 (vehicle/15 min) on testing
Khadilkar et al. [85]	2018	Scheduling Railway Lines	Real single- and multi-track railway data of different routes with various number of trains and stations in routes	Reinforcement learning	Scalable to large scale dataset due to transfer learning; manage large, realistic problem instances in computation times and outperform other traditional techniques.
Zhang et al. [86]	2020	Short-term Passenger Demand Prediction	Real taxi dataset of New York City from January 2016 to June 2018 for 63 zones	MTL-TCNN	An automatic feature selector algorithm; outperform other models with 2.5% RMSE accuracy
Cheng et al. [87]	2018	High-Speed Trains Positioning	Beijing-Shanghai high-speed railway real-data contains 725 groups of data	KNN	Improve KNN performance by applying ant colony optimization (ACO) and online learning algorithms; obtain a better cluster number of positioning data; Outperform other algorithms with 2.21 MAE accuracy.
Alawad et al. [88]	2019	Railway Safely and Accidents	Real data of accidents and passenger information like passenger age and time of accident occurrence for 71 accidents	DT	Developed a classification model regarding the occurrence of accidents with good prediction accuracy of 88.7% on test data

provides the most recent works based on ML in the field of intelligent transportation.

3.5. Load Forecasting Applications

Accurate load forecasting, both short- and long-term, is an essential task for the daily (economic) dispatching of electricity, both to prevent wasting energy production and integrating renewable energy resources.

Energy companies monitor, control, and schedule load demands and power generation to enhance energy management systems. However, electrical load profiles are becoming more complicated, not only because of the stochastic behavior of customers, but also because of the introduction of new non-linear components in power systems, such as electric vehicles, buses, and bikes. Therefore, many researchers have been developing both deterministic and probabilistic load forecasting models to improve the precision and speed of prediction models.

Table 6. Overview of Research Using Machine Learning for Load forecasting.

Reference	Year	Application	Data	Method(s)	Remark(s)/Contribution(s)
Zhang et al. [89]	2020	Medium-term Load Forecasting	Two real-world datasets: New York (1200 hourly electricity demand values of February 2018) and Queensland (1200 half-hour electric load values of January 2017) region	VMD, SR, SVR, CBCS	This study proposed a novel hybrid model based on variational mode decomposition (VMD), self-recurrent (SR) mechanism, support vector regression (SVR), chaotic mapping mechanism, and cuckoo search (CBCS). The VMD-SR-SVRCBCS outperformed other medium-range prediction methods (240 half-hours window) in both cases with 2.5 and 0.9 MAPE of New York and Queensland, respectively.
Feng et al. [90]	2020	Short-term Load Forecasting	Real hourly load data of University of Texas as Dallas for 2014 and 2015	Reinforcement learning	This study proposed a deterministic and probabilistic load prediction using the two Q-learning agents to select the best model locally from deterministic load forecasting methods and ten state-of-the-art ML-based models. The results show 50–60% accuracy improvements compared to single-phase benchmarks models.
Ahmad et al. [91]	2019	A-Day Ahead Load Forecasting in Smart Grids	Real hourly data of two USA grids (DAYTOWN, Ohio and EKPC, Kentucky) for two years (2014–2015)	ANN	This study considers both accuracy and execution time to develop their model to scale well in bigger datasets. The authors introduced the pre-preparation, prediction, and optimization modules. Taking advantage of a heuristics-based optimization method minimized MAPE while reaching 98.76% accuracy, which was relatively better than existing bi-level techniques.
Zheng et al. [92]	2017	Short- and Medium- Term Load Prediction	Real hourly data of electricity load of ISO New England (2003–2016)	PCA, LSTM, XGBoost with K-means	The authors presented a hybrid algorithm based on supervised and unsupervised machine learning techniques as follows: firstly, they applied empirical mode decomposition (EMD) and similar days selection days to extract dominant features, then, they made predictions with LSTM considering a very rich dataset for 11 years for training, one year for validation, and one year (2016) for testing. The similarity between days achieved by XGboost-based weighted k-means. The testing results for one-day and one-week ahead shows this hybrid method improved the average accuracy of the LSTM-based model from 5.43 to 1.08 MAPE and 8.74 to 1.59 for a day ahead and a week ahead, respectively.
Dabbaghjamanesh et al. [93]	2020	A-Day Ahead Load Forecasting for EV Charging Station	Synthetic dataset with hourly resolution	Reinforcement learning	This study proposed a reinforcement learning-based model to predict a day ahead EV charging station load demand. The proposed Q-learning model outperformed CNN and RNN models in three different scenarios (coordinated, uncoordinated, and smart charging) in terms of MSE metrics. Higher accuracy, higher speed, and flexibility are three main advantages of the proposed model.
Farsi et al. [94]	2021	Short to Long Term Ahead Load Forecasting (1–30 days ahead)	Real datasets of hourly load consumption of Malaysia (2009 to 2010) and Germany (2012–2016)	CNN and LSTM	This article proposed a parallel LSTM-CNN Network (PLCNet). Compared to others, this study’s main advantage is to use LSTM and CNN in parallel and concatenate their outputs with a dense layer to make the final prediction. The proposed method outperformed statistical and machine learning models with 98.23% R-square accuracy for Malaysians and improved Germany’s R-square accuracy from 83.17 to 91.18% for a day ahead load prediction.
Hafeez et al. [95]	2020	A-Day Ahead Load Forecasting	Real hourly load data of three USA power grids (FE, EKPC, and Dayton) from 2005 to 2012	ANN (restricted Boltzmann machine)	The authors introduced a hybrid model based on a deep neural network (restricted Boltzmann machine), modified mutual information (MMI) technique to extract features, and proposed a genetic wind-driven (GWDO) optimization method to adjust the model’s parameters. Together with their fine data engineering procedure, this new optimization algorithm helps to improve the MAPE accuracy between 4.7% to 17.3% compared to benchmarks. Moreover, their model’s average convergence time rate is 52 s which is less than 58–102 s of benchmarks’ expectations time.
Han et al. [96]	2019	Medium to Long Term Load Forecasting (a week to a year)	Two hourly real daily load datasets, Hangzhou from January 2015 to March 2017 and Toronto from May 2002 to July 2016.	CNN and LSTM	The authors proposed two methods, time-dependency convolutional neural network (TD-CNN) and cycle-based long short-term memory (C-LSTM), that outperformed other benchmarks in terms of accuracy and execution time. Their models’ main advantages are extraction of the long-term global combined features and short-term local similar features in the LSTM-based model and conversion of load’s temporal correlation into spatial ones in the CNN-based model.
Chen et al. [97]	2019	A-Day Ahead Load Forecasting	Two hourly real datasets North American Utility and the ISO-NE from 1985 to 1992; the datasets of 1991 and 1992 were used for testing	ANN with residual connections	This study introduced a deep neural network with residual connections, one of the well-known techniques to overcome the problem of lost information in earlier layers in a deep network by applying direct links from primary layers to deeper ones. Applying ensemble strategy on the two rich datasets provides the generalization capacity of their model. The proposed model improved the MAPE error rate from 1.48 of the best benchmark model to 1.447 in the ISO-NE dataset and from 1.73 to 1.575 for the North-American utility dataset, which also implies the robustness to temperature variation of the proposed model.
El-Hendawi et al. [98]	2020	A-Day Ahead Load Forecasting	Real dataset of the hourly electric market of Ontario, Canada from 2011 to 2016	ANN	The authors used the wavelet transform to decompose the input data into different levels with different frequencies to feed several neural networks. Instead of having one model, they trained different neural-based models with part of transformed input data and made final forecasting considering all models’ predictions. The proposed ensemble model improved the MAPE accuracy by 20% compared to other traditional neural networks.

presents recent advancements of machine learning studies in load forecasting.

4. Discussion

ML-based algorithms have shown remarkable results in power system analytics compared to traditional methods. However, even if the models proposed by the literature showed to work fine in real datasets, their performance in industrial applications has not been sufficiently demonstrated yet, due to cost or privacy issues. This suggests the need for further investigations at the industrial level, where the presence of input data with different distributions or big data properties (e.g., volume, velocity, variety, and veracity) could decrease the performance of ML models.

Regarding the data used for system validation, the studies generally presented customized datasets. They typically provided information on the total number of samples, sampling frequency, recording time, and percentage of data used for training and validation. As several datasets were synthetically generated using simulation software, only various studies reported problems with imbalanced datasets and missing items in the data. In this regard, Hong et al. [45] analyzed the case in which data were missing from one of the buses, concluding that system performance decreased significantly. Karagiannopoulos et al. [46] extrapolated historical data and used information from the public domain or from neighboring systems to deal with missing or noisy data. In this sense, Hafeez et al. [95] replaced missing values with the average values of preceding days, while El-Hendawi et al. [98] replaced missing data with the average values of the same day in previous years. Similarly, Ray et al. [75] used measurements from past hours to fill in missing data and performed data cleaning to exclude incorrect data from training. Jia [79], Ou et al. [84], and Alawad et al. [88] also highlighted the need to clean up missing data, while Li et al. [44] wrote the missing features as zero to keep the dimension of the matrix constant. Additionally, Gao et al. [73] presented an ML-based fault detection system in a photovoltaic array and quantified the impact of missing PV input data (irradiance, temperature, and different combinations of them) on system accuracy. On the other hand, Li et al. [83], Vantuch et al. [54], and Liao et al. [53] discussed the effect of the imbalanced dataset on performance. In this sense, Wang et al. [57] solved the data imbalance problem using an enhancement method that equalized the amount of data (random cropping of existing data to generate a new dataset, increase of random noise, signal reversing, etc.). Similarly, Jia [79] applied a synthetic minority over-sampling technique that addressed the dataset imbalance problem without overfitting the classifier.

The lack of standard datasets for the testing of ML-based algorithms also emerged as a relevant issue. Indeed, all the models presented in the literature are usually tested on not-standard datasets, with very different characteristics and peculiarities, thus making the comparison of the performance of such methods almost impossible. It is then apparent that, when it comes to the selection of the most suitable ML method to be implemented in large scale applications, this lack of information represents a relevant issue, that would eventually prevent the implementation of novel (and potentially more performing) methods in favor of (probably less performing) traditional ones. This, in the end, highlights the need for the definition of application-specific standard datasets, to allow a fair comparison between the very different ML methods proposed for each application. The standardized dataset should have the following properties:

Size: considering the industrial side, the dataset size should be considerably big with high dimensionality. Although some weak learners, such as DT, showed to work perfectly with a small amount of data, they would not well generalize in the big domain. On the contrary, neural network models have better accuracy results in the big domain;

Quality: if the focus is only on the performance of the machine learning model, the different input datasets should have the same properties. For example, some models are very robust to none values or outliers while others are not. Preparing a dataset before feeding it to a model relates to data engineering procedures rather than to the model performance;

Validity: the dataset should accurately represent the phenomena or reality of events. The statistical properties of the standardized dataset should be as much as possible close to real-life scenarios to show how practical models are;

Uniqueness and completeness: the information should be unique and not be duplicated over the dataset to make sure a trained model will generalize well enough in actual cases. Moreover, it should cover all the possible occurrences or conditions. When considering, for example, the power quality disturbance classification, the dataset should include all the essential distortions;

Train and test division: it is important the make sure that the performance of all models are evaluated with the same train set. Otherwise, a chosen test set probably only consists of easy instances, or it does not consist of all the possibilities;

Accuracy metrics: authors used different metrics to evaluate their model performance; however, it is not possible to compare various studies when the same accuracy metrics are not used. The metrics should be proposed taking into account the nature of problems. For example, there are much fewer abnormal events in anomaly detection than normal, so the model with 99% accuracy does not guarantee that it correctly detected all abnormal events; for such studies, F1-score or AUC should be taken into account.

Researchers proposed different models based on one or more techniques. Figure 3. shows the frequencies of techniques presented in the literature review of this study. In this figure, ANN consists of the traditional neural network such as MLP and Boltzmann machine, SVM includes both classification and regression, and PCA encompasses all PCA methods.

Alternatively, it seems the hybrid models had better performances compared to others, particularly the one that combined feature engineering techniques with prediction models. Reinforcement learning methods such as Q-learning have also enhanced accuracy in some applications like intelligent transportation systems and load forecasting. In some applications, such as PV prediction or load forecasting, which deal with temporal datasets, some sequential techniques such as GRU or LSTM are preferred.

5. Conclusions

When facing the challenges related to the management of smart power systems, it became apparent that traditional techniques are no more computationally promising solutions. One of the limitations of conventional algorithms is their inadequate capacity to handle a large amount of data—consisting of chunks of heterogeneous datasets—collecting from measurement devices such as phasor measurement units and smart meters. As a result, many researchers developed high-level, efficient, and reliable solutions based on state-of-the-art intelligent learning algorithms to provide innovative solutions or promote the overall performance of current models in various power system fields. In this context, the ML paradigm and modern ML algorithms are categorized and presented in this article. Furthermore, this study provided a systematic overview of the latest machine learning techniques and models employed to bring new resolutions in power flows, power quality events, power quality parameters, photovoltaic systems, intelligent transportation systems, and load forecasting services. The authors also suggested the properties of a standard dataset for testing and reviewing the ML-based models to make a fair comparison between the performances of proposed models for each topic. However, the literature analysis implies that hybrid models based on supervised machine learning algorithms are applied more exceeding than unsupervised or semi-supervised techniques. Thus, it can be highlighting that supervised algorithms convey more benefits to problems typically faced by electrical power engineers. Finally, it can also be concluded that the application of machine learning methods in electrical systems simplifies complex issues and ensures more reliable and accurate results. As numerous works proposed solutions based on ML techniques, the authors limited their research to well-known newly published articles. Accordingly, in future work, the authors focus on and review articles related to each topic separately to provide an informative survey.