Deep Learning-Driven Ensemble Learning for Solar PV Faults Detection Under Low Irradiance Condition

Solar photovoltaic (PV) system has become one of the most prevalent renewable energy sources, offering numerous benefits. However, such PV failures are inevitable due to circuit issues or environmental changes. This study proposes a low irradiance PV faults detection approach based on deep learning-driven ensemble learning (ELDL). Random forest, gradient boosting, and AdaBoost are ensembled as base learners while deep neural network (DNN) is used as final decision-making of meta learning. Both internal and external features are covered as they all affect the PV performance. Double dimensionality reductions of principal component analysis and linear discriminant analysis are applied sequentially to gain representative feature extraction. Different tuning strategies are adopted: 1). Hyperopt for base learners; and 2). Optuna for DNN. Hyperopt helps gain more conditional search spaces, which become difficult due to complex, different base learners in ensemble learning. As such, Hyperopt provides versatility in search spaces. Optuna gives strong generalization of DNN in optimizing hyperparameters. The proposed ELDL method is not only validated with metrics but also with a robustness test of noise injection. The results show that the ELDL method is superior to other benchmarking models.