Classification of Photovoltaic Faults Using PSO-Optimized Compact Convolutional Transformer

Diagnosing photovoltaic (PV) farms has become increasingly complex due to their large-scale presence in diverse environmental conditions. A comprehensive diagnosis process comprises four essential steps: detection, localization, classification, and remedy. This study primarily focuses on the classification of faults in grid-connected PV arrays, including line-to-line faults, open circuit faults, short circuit faults, and partial shading conditions. The deep learning-based Compact Convolutional Transformer (CCT) is employed for classifying these PV faults. To eliminate the need for heuristic parameter/hyperparameter tuning of the CCT model, this paper utilizes Particle Swarm Optimization to optimize parameters/hyperparameters such as kernel size, pooling size, stride, padding, number of multi-heads, and the number of transformer encoders. Given that CCT operates based on images, the study investigates the use of heat maps incorporating different sizes, DC/AC signals, and the number of fault signal cycles as inputs. To reduce the training dataset for CCT, Taguchi experiments are employed to generate orthogonal data while considering variations in irradiance and temperature. A realistic PV array subset is used to demonstrate the performance of the proposed method. The simulation results reveal that the proposed approach outperforms classical machine learning algorithms (Support Vector Machine, Decision Tree, K-Nearest Neighbor, and Random Forest) as well as convolutional neural network (CNN)-based models (AlexNet, Googlenet, ResNet50, VGG16, and VGG19). Specifically, the proposed method achieves the highest testing accuracy (97.34%) and ResNet50 exhibits the second best testing accuracy (93.237%) among all CNN-based the methods while Random Forest demonstrates the highest testing accuracy (84.24%) among classical machine learning methods.