A one-dimensional convolutional neural network framework for photovoltaic array fault detection

Abstract Faults in the photovoltaic (PV) array radically disrupt power production, system operation, and efficiency. Detecting the PV array faults becomes a challenging task as fault characteristics and normal operating conditions exhibit the same characteristics under certain conditions. This paper introduces an innovative framework for fault detection in PV arrays using a one-dimensional convolutional neural network (1D-CNN). Initially, the fault signatures of hot spot faults, line-ground faults, and short circuit faults in the 3.2 kW PV array are dissected to determine the appropriate measurements for fault detection and discrimination. It is found that the PV array measurements, such as voltage, current, power, and their derivatives, provide distinct fault characteristics, and it would be helpful to discriminate between the faults. The datasets used for training and testing are generated with the selected measurements under different operating conditions. The effectiveness of the proposed 1D-CNN-based PV array fault detection (1D-CNN-PVFD) framework is appraised using confusion matrix analysis and contrasted with the recently presented machine learning techniques found in the literature, namely kNN, SVM, and ANN. The 1D-CNN-PVFD framework outperforms the different techniques with an accuracy of 99.44% and demonstrates promising consistency in performance.