Adaptive Automatic Generation Control in Multi-Area Power Systems: A Deep Reinforcement Learning Approach for Model-Based Controller Tuning

The increasing penetration of renewable energy sources (RES) reduces power system inertia, causing severe frequency excursions following disturbances. Conventional Automatic Generation Control (AGC) using fixed-gain controllers struggles to adapt to these rapidly changing dynamics. While deep reinforcement learning (DRL) yields improved transient stability, reduced overshoot, and better disturbance rejection for adaptive AGC, most studies lack validation on industrial-grade platforms. This paper addresses this gap by developing a comprehensive DRL framework for optimal AGC tuning via a novel PowerFactory-Python co-simulation interface. Four DRL algorithms such as TD3, SAC, PPO, and DDPG, are benchmarked for the real-time optimization of PI, PID, and Fractional-Order PID (FOPID) controllers. Extensive simulations on a modified IEEE 14-bus system with 46% renewable penetration demonstrate that TD3 provides the most stable frequency response, preventing dangerous oscillations during critical transitions like generator restoration. The results establish a clear performance hierarchy: TD3 outperforms SAC, followed by DDPG, with PPO being the worst performing. Among the evaluated control structures (PI, PID, and FOPID), FOPID achieves the best overall damping; quantitative evaluation shows that TD3-FOPID reduces the Integral of Absolute Error (IAE) by 55.9% compared to conventionally-tuned baselines. Finally, configuration training times fall within a 24-hour window, proving compatibility with day-ahead economic dispatch scheduling for overnight retraining. This high-fidelity co-simulation framework provides a practical pathway for the industrial adoption of intelligent AGC.