3.1.7. Neural Network Learning

Neural networks, also referred to as artificial neural networks (ANN), are derived from the biological concept of brain cells called neurons. Therefore, to understand ANN, one has to be familiar with how neurons work [74]. An ANN functions on three layers (input layer, hidden layer, and output layer), in the same way as the brain neurons work on four parts (dendrites, nucleus, soma, and axon) [74]. The input layer receives the data, which is then processed by the hidden layer, before it is sent as a calculated output to the output layer [76].

#### 3.1.8. Instance-Based Learning

In this case, the algorithm learns a specific pattern and then applies it to new data [74]. This learning method becomes more sophisticated as the amount of data grows [77].

#### 3.1.9. Evolutionary Computation

Evolutionary computation is currently a distinct branch of artificial intelligence inspired by nature [78,79], with smart methods based on evolutionary algorithms targeted at solving various real life problems through natural processes involving live things [79]. It is based on random processes, data regeneration, and data replacement within a system, such as a personal computer or any other data center. A variety of evolutionary computation approaches are utilized for different applications, including image processing, cloud computing, and grid computing [79].

#### **4. Artificial Intelligence-Based Failure Detection and Diagnosis Methods**

Most of the AI-based methods used in failure detection and diagnosis adopt ML models, such as support vector machines, wavelets, neural networks, fuzzy logics, decision trees, graph-based semi-supervised learning (GBSSL), regression, etc. for the development of models and algorithms that are trained to learn the relationships between the input and output parameters of PV systems. The data are obtained from experimentally accurate PV model measurements, and then split into training and test datasets. This section reviews and discusses the AI-based detection and diagnosis techniques proposed in various literatures. Figure 1 shows a distribution of the available literature on fault detection and diagnosis methods for PV systems based on the mentioned ML models. It presents a rough estimate of the number of conference and journal papers published to date. In general, the amount of publications on the subject is relatively low, which is due to the fact that it is a newly explored area of research that dates back to only about 15 years.

**Figure 1.** Distribution of AI-based fault detection and diagnosis methods for PV systems found in the available literature.

#### *4.1. Neural Network-Based Methods*

The study presents a literature review concerning the AI methods of fault detection and diagnosis that are neural network based. Table 1 presents a summary of the different methods presented in the subsection, pointing out the main contribution of each reviewed literature.

In [80], a simple and effective Bayesian neural network (BNN) model for estimating power losses in PV plants owing to soiling was devised. Four models were built based on the Bayesian neural network (BNN) to assess the performance of two plants for dirty and clean module conditions under standard test conditions (STCs). The loss due to the soiling impact is shown by the difference in the STC power between the two conditions. The study found that utilizing a BNN model rather than a polynomial model for calculating the STC power of a PV system is more successful due to various factors that affect the polynomial model's performance, including the database size. Figure 2 presents a schematic diagram of the BNNs used for calculating the STC power, consisting of an input layer (which has solar irradiance and cell temperature as input), a single hidden layer (estimated during the training process), and an output layer (which provides the STC power output produced by the plant). Bayesian regularization, a process of updating the weight and bias values according to the Levenberg-Marquardt optimization technique, which helps in reducing a combination of squared errors and weights, and determining the correct combination to produce a network that generalizes well, can greatly improve the generalization ability of neural networks [81]. BNN is basically back propagation (BP) with an additional ridge parameter added to the objective function [80]. The study provides an important contribution as it helps the operation and maintenance personnel in decision making between washing cost and losses in energy production.

**Figure 2.** Schematic block diagram of Bayesian regularization [80].

In order to successfully detect and categorize PV array faults, the authors of [82] employed a deep two-dimensional convolutional neural network (CNN) to extract features from two-dimensional scalograms generated from the PV system data. The study took into account five different fault scenarios as well as the use of MPPT. There are two variations of the proposed method. First, the last layers of a pre-trained AlexNet CNN [83] are finetuned to generate a six-way classifier in the first configuration. In the second configuration, features are extracted from a specific layer of a pre-trained AlexNet and then combined with a classical classifier. The suggested model's performance is compared to machine learning- and deep learning-based models. The suggested method surpasses the previous methods in terms of detection accuracies for both noisy and noiseless data, according to the authors. They also illustrated the need of representative and discriminative features for categorizing errors (rather than using raw data), especially in noisy environments. Automatic feature extraction based on deep learning has been found to be superior to manual feature extraction. In order to better explain the proposed method, Figure 3 presents a flowchart showing the proposed PV array fault diagnosis method and existing methods. Another neural network-based method of fault detection and diagnosis in solar PV systems, which uses Elman neural network (ENN), is presented in [84]. The study examines the implicit mining link between original data and fault types, develops multiple hypothesis models, and analyzes the mean and variance of diagnostic errors to determine which diagnostic model is optimal. The suggested fault diagnosis approach based on ENN overcomes the problem of PV system multi-source and multi-type defect identification by minimizing the number of sensors, which only collects PV operation data and data from the atmospheric environment.

The main contribution of [66] is the proposition of a technique for isolating and identifying faults that occur in the PV system, and its implementation into a field programmable gate array (FPGA) with real-life application effects. The proposed approach detects and diagnoses faults that occur in PV bypass diodes, cells, modules, and strings. It accomplishes this by examining a set of parameters, such as current, voltage, and the number of peaks in the I-V characteristics that indicate normal and abnormal PV system operation. Two separate algorithms are used in this strategy. The first algorithm isolates defects with different combination attributes using a signal threshold approach. The second technique uses an ANN-based approach to identify errors that have the same mix of features. The approach is low-cost and easily adaptable to large-scale PV systems. The block diagram of the proposed fault detection technique based on the threshold approach and ANN is presented in Figure 4.

**Figure 3.** Flowchart showing the proposed method of PV array fault diagnosis and existing methods [82].

**Figure 4.** Block diagram of the proposed fault detection technique based on the threshold approach and ANN [66].

A method for detecting the islanding phenomenon in the PV system was introduced in [85]. To detect islanding actions, the method employs a multi-variable method based on an extended neural network that combines passive and active detection modes. The method combines the extension theory's extension distance with a neural network's learning, recalling, generalization, and parallel computing capabilities. The study used an extension neural network to distinguish between power quality interference (voltage swells, voltage dips, power harmonics, and voltage flickers) and actual islanding operations at the grid power end, in order for the islanding phenomenon detection system to cut off the load correctly and promptly when a real islanding operation occurs. The detection algorithm is created and translated using a PSIM software package based on the C language and written in dynamic-link library (DLL) modules. The signals sent by DLL are passed back to the controller to complete the islanding detection control. An enhanced machine learning based approach for the detection and diagnosis of short-circuit faults, and a complete disconnection of the string from an array in the DC side of the grid-connected PV system is

presented in [86]. The process uses a probabilistic neural network (PNN) classifier with one diode model (ODM) and a parameter extraction method to create a trustworthy model of a real-world PV system. Two PNN classifiers are used in the proposed method, one is for detecting fault occurrences and the other for diagnosing the type of fault. There are four stages to the method: Array parameter extraction; experimental model validation; elaboration of database of both healthy and problematic operations; and network design, training, and testing based on the best-so-far ABC algorithm. The contribution of this study is highlighted in the model's ability to detect a fault, while also pinpointing its origin. However, for the method to be effective, the high-quality database which is not always readily available, is required to deal with classification problems. To deal with this issue, the authors suggested having in place a trusted simulation model, which is able to mimic the exact healthy and faulty behaviors of a PV system. Another PNNbased intelligent method for PV system health monitoring was proposed in [56], which can detect and categorize short- and open-circuit faults in real time, as well as locate the faulted PV string in a grid-connected PV system. To detect and diagnose faults, the suggested technique uses data obtained from various sensors in PV systems, such as voltage, current, irradiation, and temperature which are used to deliver information on fault occurrence. The PNN used in the method has four layers: The input layer (the number of neurons in the layer represents the number of training and testing samples); the hidden layer (whose pattern units are equal to the training set sample space); the summation layer (the number of neurons is equal to the number of sample space classes); and the output layer or the decision layer (containing one neuron which provides the classification decision). Moreover, it was developed and validated in computer programs utilizing a novel approach to PV system modeling that only requires data from the manufacturer's datasheets provided under normal operating cell temperature conditions (NOCT) and STCs. The modeling approach is an improvement to the previous approaches where STC conditions, I-V characteristics or NOCT conditions are used but never combined together. This systematically builds a relationship between the ideality factor, thermal voltage, and series resistance with the PV module temperature using the manufacturers' datasheet elements. The PV system simulation model is then used to implement and validate the PNN-based detection model and classification method. The authors of [87] explored real time online fault detection for PV modules under partial shading conditions. The approach suggested in the paper is an intelligent method that uses artificial neural networks (ANN) to estimate the output PV current and voltage under varying operating conditions utilizing solar irradiance and cell temperature meteorological factors. Since it performs the real time correlation of estimated performances with the measured performances under variable conditions, the method can also be used to detect the possible anomalies in PV modules. The model proposed is independent of the measured PV module performance which makes its system of fault detection autonomous. Figure 5 depicts the fault detection flowchart of the proposed method. The results of the study show that the proposed method can accurately estimate the output and detect any decrease in output power without requiring any complex calculations or mathematical models. However, it does necessitate that the ANN be trained on a regular basis in order to accurately estimate the output parameters. The approach could also be used in PV arrays or large-scale PV plants, as well as in low-cost microcontrollers for real-time applications.

**Figure 5.** Proposed fault detection flowchart diagram [87].

According to the authors in [88], few data collecting systems in solar power plants focus on intelligent reasoning of the plant's state, despite the fact that the data of these collecting systems offer a wide range of capabilities. To tackle this problem, the authors presented a description of a novel data acquisition system. A Bayesian belief network-based (BNN) fault detection and diagnostic system was then built, which analyzes the acquired data for the existence of faults and intelligent reasons for potential causes of the detected faults. The BNN-based model uses a graphical representation of a problem in the form of a hierarchical network, with nodes representing random variables and directed arcs expressing the probability regarding the dependencies between these variables. Each node has a set of states, each with a probability distribution associated with it. Each arc reflects a conditional probability based on the preceding nodes. A node's state could be as simple as true or false in the most basic scenario. In a more complicated instance, the set of states could include multiple discrete states, such as low, medium, high, and very high. The measurements obtained by a BBN-based inference engine produce a change in the probability values in the respective nodes, which impacts the connected nodes in the network and leads to the automatic derivation of a decision on the likely reason of a failure. By developing a framework for analyzing sensor results and translating them to a Bayesian belief network using the Netica API, the system lays the groundwork for future advancements. The Fault Identification for Nasa Exploration Missions and Navigation (FINEMAN) system was developed as an integrated package with four main components: A connectivity interface for remotely retrieving data from the plant's data acquisition system; a preprocessor for relevant measurement selection; a fault injector for failure testing simulation; and a Netica implemented BBN interface engine. The research adopted in [64] was also ANN-based, and unlike in [87], where ANN was used for the PV module under partial shading conditions, the approach in [64] utilized AI technology for automatic MPPT fault detection and failure type judgement. The method requires five features (solar irradiation, installed capacity, MPPT power, MPPT voltage, and MPPT current) for machine learning. Each inverter's MPPT performance is gathered and modeled using a machine learning algorithm rather than a rule-based programming approach. In addition, an inefficient MPPT is discovered by comparing the real and expected power output. Moreover, faulty equipment is automatically recognized, a knowledge-based system determines the type of failure, and an alarm with the failure diagnostic is communicated to the user by a mobile device or email. The field sites collect data on power production every 5 min. As a result, the entire computation and communication process in the system takes only 5 min to complete. Using this AI

method, the authors were able to efficiently manage hundreds of projects at the same time, while also optimizing O&M performance with minimal work and resources. A novel algorithm utilizing the genetic algorithm to optimize the topology of ANN was introduced in [72]. The new algorithm is an online optimized neural network-based fault diagnostic and repairing system, aimed at providing a solution to the problem of complexity and high costs, associated with the other fault detection methods. The method offers the following factors: High speed diagnostic process as it diagnoses multiple faults in parallel; can remotely replace faulty components with good ones; can be used for modern complex PV systems; and has the ability to partition the PV panel into sub-areas. As a result, the diagnostic procedure is divided into two parts. The first is concerned with identifying the failure where the proposed method diagnoses a PV system using a genetic algorithm to optimize the topology structure of the neural network. The system is implemented in five steps: (1) Chromosome representation, which uses binary digits for topology network representation; (2) initial population, which is constructed from random individual sizes; (3) cross operator, which combines the parents to obtain two offspring using the uniform crossover; (4) mutation operation, where random mutation is used for genetic algorithm; and (5) fitness function, which is used to minimize the error value of the neural network representing the fitness function of the genetic algorithm. The second part of the proposed method is solely concerned with determining the cause of the failed area by dividing the structure of the PV system into three modules (the PV panel module, the charger module, and the battery module). Following this step, the ANN is trained with multiple types of faults for the three divided PV system structures and then the ANN begins the diagnostic process. When the proposed technique's results were compared to fuzzy-based and classic neural network-based diagnostic systems, it was discovered that the proposed method produced better results. The authors of [89] presented a neural network-based method for modeling the relationship (MPP) of a shaded PV array and environmental parameters (solar irradiance, sun angle, ambient temperature) in non-uniform settings. Similar to the shading factor, this neural-network-based function can characterize the shadow impacts on a solar PV array over time. As a result, the neural network model is able to eliminate the inaccuracy produced by the shading factor's complex calculation. In contrast to the prior efforts that only address the uniform shadow on the PV array, this method considers the non-uniform shadow and illumination. The proposed method's procedure is as follows:


The authors of [90] proposed a PV prognostics and health management (PHM) technique. The system was created to track the health of photovoltaic systems, measure degeneration, and provide maintenance recommendations. It employs a system-specific ANN model, which eliminates the need for prior knowledge of system components and design. Two detection techniques were tested in order to better monitor the health of the PV system. The energy loss fault detection system uses a neural network model to compare the sum of power loss over a lengthy period of time. An alarm threshold can be set to detect the long-term effects, such as soiling or material degradation, and alert the user to the need for maintenance. In the event of a catastrophic system failure, such as the loss of a

string of modules or an inverter failure, the acute fault detection technique examines the potential of the PV system performing below model expectations and should warn the user. When the two techniques are combined, the short- and long-term PV system defects can be detected. The metrics of the two combined methods can also be used for pre-emptive inspection and maintenance, as they allow the system operator to identify the PV system failure precursors linked to failure modes.


**Table 1.** Summary of neural network-based methods.
