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

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Renewable Energy Systems

Background:

• Islanding detection is crucial for grid-connected photovoltaic (PV) systems to ensure power quality, safety, and stability.
• Conventional Active Frequency Drift (AFD) methods have limitations, including a large non-detection zone (NDZ) and fixed parameters unsuitable for dynamic grid conditions.
• A significant research gap exists in adaptive perturbation parameter adjustment for AFD methods in response to real-time grid dynamics.

Purpose of the Study:

• To develop a novel AI-driven adaptive AFD method for PV systems.
• To eliminate the NDZ and enhance system stability by dynamically optimizing perturbation parameters.
• To address the limitations of existing AFD methods in adapting to changing grid and load conditions.

Main Methods:

• Implemented a Reinforcement Learning (RL) approach to optimize the chopping fraction (Cf) and an enhanced correction factor (Cr').
• The RL agent was trained using a reward-based strategy for fast and accurate islanding detection.
• The Cr' was adaptively updated based on the rate of change of frequency (df/dt) and Cf to minimize NDZ.

Main Results:

• Achieved islanding detection times of 0.12-0.17 s, a significant improvement over standard AFD (0.2-0.5 s).
• Reduced the NDZ to below 1%, compared to 10-15% for conventional AFD methods.
• Maintained total harmonic distortion (THD) within ≤ 2%, ensuring high power quality.

Conclusions:

• The proposed AI-driven adaptive AFD method effectively eliminates the NDZ and enhances detection speed and robustness in PV systems.
• Experimental validation across various PV configurations confirms the method's scalability and reliability.
• This technique offers a promising solution for smart grids, ensuring compliance with IEEE Std. 929 islanding detection requirements while maintaining power quality and stability.