Efficient Deep-Reinforcement Learning for Photovoltaic Systems Under Faults Based on the I-V Curve Approach

Abstract

Cleaner and sustainable Photovoltaic (PV) systems need to be supervised and monitored to reduce waste energy and improve power efficiency. The proposed technique in this work enhances solar energy production by precise fault detection of short-circuit and partial shading. It extends the PV system lifespan by mitigation component and further premature replacements. Moreover, automatic fault diagnosis helps maintain steady performance in variable climatic conditions and under varying occurred faults that minimize the backup to generators and energy losses. Firstly, we introduce a Bonobo Optimization Algorithm (BOA) that is capable of extracting and identifying the unknown parameters of the PV cell to model our study PV system and to mimic the fault behaviors. The identified model is validated and then used to generate the I-V and P-V curves, which are then fed to three autoencoders (AE) within an unsupervised learning framework to extract their features. Afterward, reinforcement learning (RL) is integrated through a stacked autoencoder (SAE) to combine environmental attributes such as solar irradiance and temperature with electrical features to improve the learned features and their sparsity. Also, to enable the system to adapt dynamically to new fault scenarios and noisy environments, deep-reinforcement learning (DRL) improves feature representation and classification through Artificial Neural Networks (ANN). This methodology provides an identification and categorization of 12 selected fault types in separated and combined ways, where this technique has been applied to a PV plant located in Algeria. The classification results exhibited exceptional accuracy, achieving 100% in the training phase and 99.8% in the testing phase, even amongst noisy input conditions with 97.2%. This study provides valuable insights into improving the reliability and efficiency of PV systems, particularly in the smart IV diagnosis that used multi-string PV inverter.