Two-stage deep Q-network reinforcement learning based ultra-efficient fault diagnosis and severity assessment scheme for photovoltaic protection

Early detection of faults in photovoltaic (PV) arrays has always been the center of attention to maintain system efficiency and reliability. However, conventional protection devices have shown various deficiencies, especially when dealing with less severe faults. Hence, artificial intelligence (AI) models, specifically machine learning (ML) have complemented the conventional protection devices to compensate for their limitations. Despite their obvious advantages, ML models have also shown several shortcomings, such as (i) most of them relied on a massive amount of training dataset to provide a fairly satisfying accuracy, (ii) not many of them were able to detect less severe faults, and (iii) those which were able to detect less severe faults could not produce high accuracy. To this end, the present paper proposes a state-of-the-art deep reinforcement learning (DRL) model based on deep Q-network (DQN) to overcome all the existing challenges in previous ML models for PV arrays fault detection and diagnosis. The model carries out a two-stage process employing two DQN-based agents which is not only able to accurately detect and classify (first stage) various faults in PV arrays, but it is also able to assess the severity of line-to-line (LL) and line-to-ground (LG) faults (second stage) in PV arrays using only a small training dataset. The training and testing datasets include several voltage and current values on PV array current-voltage (I-V) characteristic curve which is extracted using the variable load technique for PV array I-V curve extraction. The model has been implemented on an experimental standalone PV array and the results show outstanding accuracies of 98.61 % and 100 % when it is verified by testing datasets in the first and the second stage, respectively.