Adaptive PV array reconfiguration using double deep q-networks for performance improvement under partial shading conditions

Partial Shading Conditions (PSCs) significantly reduce the generation efficiency of PhotoVoltaic (PV) systems. PV Array Reconfiguration (PVAR) is an effective approach to mitigate PSCs. However, the existing methods tend to have high computational complexity, require re-optimization, and lack robustness against dynamically changing shading patterns. This paper proposes an adaptive PVAR based on a Double Deep Q-Network (DDQN). The PVAR employs controllable switches to enable multiple feasible electrical configurations, while the DDQN learns to select effective switching configurations to maximize power output under diverse synthetic PSC scenarios. The results show that by mitigating the bias of DQN and stabilizing learning, the DDQN outperforms all baseline methods (fixed series, parallel, series-parallel configurations, logic-based method, and the DQN). It achieves the highest energy yield, the lowest average Mismatch Loss (ML), and the highest Fill Factor (FF), while maintaining low inference latency. In addition, a switching-aware evaluation shows that unnecessary switching operations can be substantially reduced with only a limited reduction in energy yield. Consequently, this study establishes an intelligent and computationally efficient framework for adaptive PVAR under diverse synthetic PSCs and highlights its potential for practical implementation.