Convolutional Neural Network with Oscillating Activation Functions for Fault Detection and Classification in Photovoltaic Arrays Using Power–Voltage Curves

Effective fault monitoring and diagnosis in photovoltaic (PV) systems are essential for optimizing their performance. This study presents a novel approach, using one-dimensional convolutional neural networks (1D-CNNs) equipped with innovative oscillatory activation functions. These novel oscillatory activation functions, unlike traditional Rectified Linear Unit (ReLU) function, exhibit non-monotonic behavior, which mitigates the “dying neurons” problem by preventing the network from becoming stuck in a state where a significant portion of neurons consistently output zero. This key innovation enables the model to more effectively capture the non-linear dynamics within the PV system, leading to improved fault classification accuracy. Based on the analysis of differences in the P–V curves of PV panels under different fault states, the P–V curves, temperatures, and irradiances are taken as input data. The model integrates an oscillatory activation function into a 4-layer structure after Batch normalization. This approach demonstrates a significant improvement in detection and classification accuracy, exceeding 99% for various PV faults, such as short circuits and partial shading. Furthermore, the use of oscillatory activation functions contributed to a more compact network architecture, reducing the number of neurons required and allowing for rapid convergence during training to achieve this level of performance. This improvement is notable compared to traditional methods that rely on the ReLU activation function. These promising results pave the way for more efficient management and maintenance of PV systems, thus promoting a more effective energy transition.