Diagnosis of PV faults using digital twin and convolutional mixer with LoRa notification system

Abstract

The amount of energy that is generated using photovoltaic (PV) arrays has been increasing rapidly over recent years. To avoid energy and financial losses due to PV faults, numerous methods for diagnosing PV faults have been proposed. Both digital twin (DT) and deep learning (DL) have been proven to be effective in solving detection/classification problems in various fields. This work develops a method of PV fault diagnosis that has the following three stages: (1) Detection of faults in a PV array using DT, (2) classification of the detected faults using ConvMixer, and (3) notification of detected and classified faults using a LoRa (long-range) system. DT is a virtual/digital model that is designed to reflect accurately the behavior and characteristics of a physical object. This work implements the model-based DC power generation of PV arrays using DT. In this study, a new DL method, called the convolutional mixer (ConvMixer) based on patch embedding and combining depthwise and pointwise convolutions to classify PV faults, is presented. The inputs of ConvMixer are 2D images generated from data on PV DC array power using a Markov transition field (MTF) transform. A LoRa notification system is very suitable for use in low-power wide-area networks (LPWANs), such as those needed in large PV farms. Simulation results demonstrate that the proposed ConvMixer outperforms other classical machine learning (ML) methods, such as decision tree, k-nearest neighbor, random forest, and support vector machine methods, as well as other classical CNN-based methods, such as AlexNet, ResNet50, VGG16, and VGG19. A real-time digital simulator (Opal-RT eMegasim) is used to verify the real-time applicability of the integration of DT, ConvMixer, and the LoRa notification system.