Solar photovoltaic array short circuit fault analysis with machine learning using pre-trained convolutional neural network for feature selection

Abstract

Solar photovoltaic (PV) array deployment is rapidly increasing, but faults, particularly short-circuit faults, pose significant reliability and safety challenges. Machine and deep learning techniques have been applied to accurately identify PV short circuit faults using current–voltage (I–V) characteristic curve data; however, traditional machine learning models often require manual feature selection, and deep learning models demand large datasets, which are challenging to obtain. Additionally, shading effects are typically not included as operating conditions. This paper, therefore, proposes a machine learning-based fault detection and classification (FDC) method using pre-trained convolutional neural networks (a type of deep learning) for automatic and efficient feature selection, aiming to maintain high accuracy, fast FDC time and low memory usage while requiring less training data. The training and testing I–V curve data were generated from a PV array modeled in detail in MATLAB/Simulink. Furthermore, faults were simulated under varying irradiance, mismatch levels, and shading effects. The evaluated pre-trained convolutional neural networks include AlexNet, VGG, GoogleNet, ResNet, SqueezeNet, DenseNet, ShuffleNet, and EfficientNet. Among these, EfficientNet paired with a support vector machine demonstrated the best performance, achieving over 95.5 % across all performance metrics, with an FDC time of 4.23 s and a feature selection stage memory usage of only 20 MB. This approach can be integrated into PV system health monitoring to facilitate early FDC, enhancing system lifetime, safety, and reliability.