Accurate detection of critical LLFs and LGFs in PV arrays based on deep reinforcement learning using proximal policy optimization (PPO)

Critical line-to-line faults (LLFs) and line-to-ground faults (LGFs) in photovoltaic (PV) systems are the most difficult faults to detect not only by conventional protection devices, but also modern fault detection schemes. The difficulty occurs due to critical mismatch levels and/or high fault impedance values which result in LLFs and LGFs remain undetected thus damaging the PV components, affecting system stability, reliability, and efficiency, and even leading to catastrophic fire hazards. However, challenges persist even in recent studies, including the need for a massive training dataset, disregard of fault severity assessment, and insufficient model accuracy. To address these challenges, the present paper proposes a deep reinforcement learning (DRL)-based model to detect, classify, and assess the severity of all and specifically critical LLFs and LGFs in PV arrays using the proximal policy optimization (PPO) algorithm. Additionally, to carry out the dataset dimensionality reduction, thus simplifying the training process, a two-stage feature engineering process has been implemented, including a feature importance finding stage using the permutation technique and a feature selection stage. To implement the proposed model and verify its capability in real-life condition, a laboratory PV system has been carefully designed. The results of the real-world experiment shows that the proposed model is able to detect LLFs and LGFs, under various environmental (temperature and irradiance), and electrical (mismatch and impedance) conditions with outstanding 100% of accuracy in the test process, using only a small training dataset.