A deep reinforcement learning-based adaptive multivariate state estimation fault diagnosis method for a photovoltaic storage grid-connected system under a high proportion of new energy integration

With the increasing penetration of photovoltaic and other renewable energy sources in power systems, the operational characteristics of photovoltaic storage grid-connected systems have become increasingly complex. The stochastic nature of renewable energy output and the alternating charging/discharging states of energy storage systems due to diurnal cycles lead to significant time-varying operational conditions, under which the performance of conventional multivariate state estimation methods for fault diagnosis markedly deteriorates. To address this, targeting high-penetration renewable energy integration scenarios, this paper proposes an adaptive state estimation-based fault diagnosis method for photovoltaic-storage grid-connected systems using deep reinforcement learning. Within the deep reinforcement learning framework, an agent is deployed to adaptively match and update modeling samples according to the real-time system state, thereby enhancing the adaptability of state estimation and the fault diagnosis accuracy. Test results under multivariate typical operational scenarios demonstrate that, compared to existing advanced fault diagnosis methods, the proposed approach achieves an average improvement of 10% in fault detection rate and reduces the missed detection rate by 3%, validating its superior adaptability and diagnostic performance in systems with complex time-varying characteristics. This research provides a new technical pathway for ensuring the safe and stable operation of photovoltaic-storage grid-connected systems under high-penetration renewable energy integration, contributing to enhanced intelligent operation and maintenance capabilities and improved fault response.