Reinforcement learning driven adaptive active frequency drift for fast and reliable islanding detection

Islanding detection remains a critical challenge in grid-connected photovoltaic (PV) systems, as failure to detect islanding conditions can compromise power quality, safety, and system stability. Existing active frequency drift (AFD) methods suffer from two major unresolved limitations: a large non-detection zone (NDZ) and fixed perturbation parameters that cannot adapt to changing grid and load conditions. Despite extensive studies on AFD, no existing method adaptively adjusts its perturbation parameters in response to grid dynamics, creating a significant research gap. To address this specific gap, this work introduces a novel AI-driven adaptive AFD method that uses reinforcement learning to dynamically optimize perturbation behavior to eliminate the NDZ and enhance stability. The proposed method employs Reinforcement Learning (RL) to optimize the chopping fraction (Cf) and an enhanced correction factor (Cr'), which is updated based on the rate of change of frequency (df/dt) and Cf to minimize NDZ while maintaining grid stability. The RL agent is trained using a reward-based learning strategy to achieve fast, accurate islanding detection while minimizing unnecessary disturbances. Simulation and experimental results show that the proposed adaptive AFD achieves islanding detection within 0.12-0.17 s (compared with 0.2-0.5 s for standard AFD) and reduces the NDZ to below 1% (compared with 10-15% for conventional AFD), while keeping total harmonic distortion (THD) within ≤ 2%. These quantified improvements demonstrate substantial gains in detection speed, robustness, and power quality. Experimental validation across different PV configurations confirms the scalability and reliability of the method. The proposed technique, therefore, offers a promising solution for modern smart grids by ensuring compliance with IEEE Std. 929 islanding detection requirements while maintaining high power quality and system stability.