Static reconfiguration techniques for PV under the corn canopy with deep reinforcement learning

Valve integrated photovoltaic (PV) systems deployed in agricultural environments are prone to power mismatch caused by shading from maize canopies. To mitigate this issue, a static reconfiguration strategy based on shadow informed deep reinforcement learning is proposed. The method collects physically consistent irradiance information under maize canopy shading and constructs a deep reinforcement learning framework in which the average power loss rate is used as the optimization objective. This framework enables offline optimization of photovoltaic cell array layouts. The results demonstrate that the array configuration discovered by the proposed framework, which learns current mismatch mechanisms and applies a global trade off strategy, achieves an average maximum output power of 1.401 W. This performance significantly exceeds that of the conventional total cross-tied topology, which produces 1.057 W, and the Sudoku-based (SDK) method, which achieves 1.103 W. The corresponding relative improvements are 32.6% and 27.0%, respectively. In addition, the static reconfiguration matrix designed using this approach achieves an average maximum power loss rate of 42.31%. This study provides a new reference for the static reconfiguration of micro-photovoltaic arrays under shading conditions and confirms the effectiveness of physics guided, data driven optimization for reliable photovoltaic system design in complex natural environments.