Improved intelligent method for detection and classification of dc series arc fault in photovoltaic system

Abstract

Series Arc Fault (SAF) is a failure that occurs between two electrical contacts and electrical circuitry. However, it is considered one of the common malfunctions of photovoltaic (PV) systems that causes serious problems, including fires and electric shocks. There are several reasons that can cause this type of failure, including incorrect installation, irregular maintenance, and environmental factors. The process of SAF detection and diagnosis is considered a significant problem as many plants with a substantial increase in their capacities are continuously coming into existence. However, to achieve safe maintenance, reliability, and productivity of large-scale PV plants, it is essential to develop a new intelligent method that presents a precise automatic detection and protection of any maloperation among thousands of PV modules. In this research, the characteristic and the behaviour of the DC series arc fault signals are analysed and modelled; nine models with different properties regarding each model are simulated. In addition, an intelligent detection and classification method that can precisely detect and classify the DC series arc fault in the PV system among the other normal or abnormal conditions are developed. Also, a validation to achieve all the requirements and further improve in the efficiency of the proposed method are presented through a comprehensive comparison with the previous methods based Artificial Intelligence including Artificial Neural Network (ANN), SVM, Fuzzy, HMM, and Convolutional Neural Network (CNN). The comparison is carried out in terms of high accuracy, fault classification ability, reliability, safety, and the Computational Complexity/Effort of the power plants. Two systems are designed and built, where each system has two levels. A change detection approach is developed in the first system using ANN which incorporates four different models with various dimensions of CNN to classify the input signal. In the second system, a Multilayer Perceptron (MLP) is used to detect abnormal signals, while a Bi-Directional Long Short-term Memory (MLP-BiLSTM) is developed to classify abnormal signals precisely. The presented systems can distinguish between different cases of signal input, including normal (inverter start-up and load change), short circuit fault, and SAF. Furthermore, various models of DC series arc fault alongside with the practical experimental records are employed with PSCAD as a tool for creating these models. Python code is used to build and evaluate the performance of the proposed methods. The performance evaluation of the two proposed systems is carried out by considering several scenarios, where each system has its own features, such as removing the vanishing, dropout problems and ensuring reliability. The achieved accuracy is approximately 98% using the proposed systems. The results demonstrated that the proposed systems have the ability to detect the series arc with a high accuracy and outperform the existing works.