Currently, the most accurate, most efficient, and most powerful machine for performing operations is the human brain, which can provide solutions to problems that PCs are not capable of solving. Researchers and scientists have developed artificial intelligence (AI) models to reproduce, to some extent, the processes that take place in the human brain [1]. Currently, AI is divided into different groups: artificial neural networks (ANNs) and different hybrid systems. Among them, ANNs are the best method as they are accurate, fast, and simple and have the ability to model a multivariate system [2].

The neural network (NN) concept has more than half a century of history; however, it is only in the last 20 years that the largest number of applications have been developed in the fields of defense, engineering, mathematics, economics, medicine, meteorology, and many others.

The history of neural networks dates to the 1940s. It was Warren McCulloch and Walter Pitts who first built a very simple neural network using electrical circuits [3]. Later, Donald Hebb proposed that neural pathways strengthen with each use, an important concept in human learning [4]. Then, in the 1950s, Nathaniel Rochester of IBM Research Laboratories first attempted to simulate complex neural networks [5]. In 1959, Bernard Widrow and Marcian Hoff developed models called “ADALINE” and “MADALINE” [6,7]. After the publication of the book “Perceptrons” by Marvin Minsky and Seymour Papert in 1969, there was a period of slowdown in research. This book argued that the concept of a single perception approach to neural networks did not have an effective correlation in multilayer neural networks [8]. In the 1970s, two competing models emerged in the conception of neural networks, called symbolism and connectionism [9]. The controversy ended with the acceptance of the symbolic paradigm as the most viable line of research. In the early 1980s, however, connectionism resurfaced, based on Werbos’s 1974 studies. These studies made it possible to rapidly develop the formation of multilayer neural networks using the so-called “backpropagation” algorithm [10,11]. Since then, the field of neural networks has seen significant advances. Some of these advances were the introduction and development of max-pooling in three-dimensional data recognition [12]. On the other hand, advances included the development of deep learning and its application to a wide variety of fields such as renewable energy [12].

However, ANNs are not the only ones that learn by example. There are other methods, such as the following. Supervised learning: this method trains algorithms on the basis of sample input and output data labeled by humans [13]; deep learning: it uses neural networks to learn from the data and to improve performance by increasing the number of samples that are available during the learning process [14]; machine learning paradigms for unsupervised classification such as conceptual clustering [15], which is an unsupervised learning method that focuses on generating concept descriptions for the generated classes. Other machine learning paradigms that learn by example are semi-supervised learning [16], active learning [17], transfer learning [18], and online learning [19].

The data collection needed to train ANNs must be a sufficiently complete and consistent set of information [20]. The development of machine learning models requires historical data from several years, supplemented by information that is more recent. This information amounts to thousands of data points [20]. In the context of different energy transition scenarios and geographical locations, it is essential to ensure that the data collection is as complete, impartial, and representative as possible. This is achieved by managing diverse and reliable sources of information. Some of the most used sources are public repositories, data from official agencies and organizations, research centers, and geographic databases [21,22]. To ensure that the data are impartial, complete, and representative of all the energy transition scenarios analyzed, it is necessary to perform careful data selection and apply measures such as domain adaptation and data augmentation [23]. Model performance can also be improved, and training data can be augmented by using pretrained models in other domains [23]. The authors, based on the objective pursued and the exact geographical location, have analyzed all data obtained in the various studies, with latitude and longitude coordinates provided in many of them. Similarly, much of the data provided are based on measurements made by the authors themselves when using AI.

ANNs have gained momentum to the point where they have become popular and useful models for classification, clustering, recognition, and prediction in a wide variety of applications [24]. ANNs are increasingly being used for different applications due to their ability and effectiveness in solving different problems. They have proven to be very efficient when it is complex to cull through a mass of existing data, for example, in the evaluation of public transportation of people and goods [25], image recognition [26], medical analysis [27], efficiency analysis in nonlinear contexts, or to adjust production functions, among other applications [28,29].

ANN consists, in most cases, of an input layer, at least one hidden layer (in the case of a simplified model), an output layer, the weight, the connection biases, the activation function, and the sum node. The layers in turn are made up of several connected units (called neurons) [30], considered to be the fundamental building blocks for the correct functioning of a neural network. The link between neurons is achieved by so-called connecting links [2]. The basic diagram of a neuron is shown in Figure 1 [31].

The main characteristic feature of ANNs compared to other approaches is their ability to learn by example. ANNs can be applied to any situation where there is a relationship between input and output variables [32]. The procedure for the learning process is what is known as a learning algorithm, the purpose of which is to alter the synaptic weights of the networks in order to achieve a previously set goal [33].

ANNs must be trained by feeding the network a set of quantified data to achieve the desired output using a pool of input data [34]. The learning process continues until the NN output matches the expected output [35]. The problem with ANN models lies precisely in overtraining, i.e., when the network capacity for training is too high or too many training iterations are allowed per network [36]. The degree of training accuracy obtained in the different applications where the ANN technique is used is very high, in the order of 10⁻⁵ to complete the training processes [2]. NNs can be grouped into different categories depending on their structure [37]. This classification is shown in Figure 2. The most commonly used are single-layer feed-forward networks, multilayer feed-forward networks, radial basis networks, and dynamic (differential) or recurrent neural networks. Of these, single-layer power supply networks are the best known and most widely used. Single-layer power supply networks were the first and simplest networks devised. Information travels in only one direction: from input nodes, through hidden nodes, to output nodes. This type of NN can be designed based on different unities, and among them, the perceptron is the most famous and simplest example [38]. Rosenblatt created the perceptron in 1958, thanks to the creation of the training algorithm [39]. The perceptron is composed of a single neuron with adjustable synaptic weights and thresholds [40]. The most frequently used algorithm is the so-called backpropagation (BP) algorithm [41]. The BP algorithm consists of training and correcting the weights until the error function is below the desired tolerance limit [37].

2. Applications of ANNs to Renewable Energies

2.1. Applications of ANNs for Wind Power and Speed Prediction

• ANN type: from the 23 references analyzed, the MLP network has been used in 17 of them, followed by the RBF in two of them. Five of them did not specify the type of ANN used.
• Structure of the ANN: the predominant type is simple with one hidden layer (70%) and the rest with two hidden layers (26%) with the exception of the investigation of [41], which uses three hidden layers. The number of neurons in the hidden layer is usually around 15, while in other cases more than 63 are selected [44].
• Amount of data: the percentage of research that makes use of data for validation is 8.69%.
• I/O configuration: the inputs to the models usually take in situ measured features such as past wind speeds, temperature, relative humidity, altitude, month, or pressure.
• Activation function: only 13 of the 23 cases detail the activation function used. In the hidden layer, linear functions are used, with tansig and logsig being the most commonly used, while in the output layer, linear functions of the purelin type are adopted.

2.2. Applications of ANNs for Solar Energy Systems

• ANN type: in most of the investigations, the MLP network has been used (24 out of 29 cases) followed by the RBF.
• Structure of the ANN: most studies use simple structures with a single hidden layer (96%), and the remaining with two hidden layers. The number of neurons in the hidden layer is usually in the order of 10, reaching 50 neurons in the research of [116]. In some cases, the number of neurons in the hidden layer is not specified, as in [104,105].
• Amount of data: the percentage of research that make use of data for validation is 6.9%.
• I/O configuration: altitude, latitude, longitude, relative humidity, or month of the year are used as the most common inputs.
• Activation function: only 19 of the 29 investigations detail the activation function used. In the hidden layer, linear functions are used, with tansig and logsig being the most commonly used, while in the output layer, linear functions of the purelin type are adopted.

2.3. Applications of ANNs for Wave Prediction

• ANN type: as in previous cases, the MLP has been the structure chosen by most researchers (17 out of 20 cases). Research using the DNN [124] and RBF [133] has also been found.
• Structure of the ANN: most of the studies analyzed use simple structures with a single hidden layer. Research has also been found that uses two hidden layers or even the research of [124], which uses three. The number of neurons in the hidden layer is usually in the order of 10, reaching 300 neurons in the research of [125].
• Amount of data: in most research, the volume of data is in the order of hundreds or thousands. Normally, a major part of the data is used for training, with the remainder applied to testing. The percentage of studies that make use of data for validation is 5%.
• I/O configuration: temperature, wind speed, wind direction, and historical wave data are normally used as inputs. Outputs predict wave heights from one hour to 24 h in advance.
• Activation function: the activation function is specified in 15 out of the 20 research. In the hidden layer, linear functions are used, with tansig and logsig being the most commonly used, while in the output layer, linear functions of the purelin and sigmoid types are adopted.