Reactive Power Optimization of Wind-PV-ESS Systems Based on Multi Agent Deep Reinforcement Learning

This paper proposes a real-time reactive power optimization strategy for distribution networks with high renewable penetration using multi-agent deep reinforcement learning (MADDPG). The strategy coordinates wind, PV, and ESS controls to minimize system losses and voltage deviations while ensuring power balance and meeting operational constraints. The MADDPG framework enables minute-level dispatch by coordinating reactive power outputs from PV inverters, wind converters, and SVCs through a methodology that centralizes learning but decentralizes operational deployment. This model-free method uses neural networks to approximate control policies, addressing convergence issues in traditional reinforcement learning. Simulations on a modified IEEE 33-node system show significant improvements, with power losses reduced by 33.10%, 38.05%, and 52.61% compared to DDPG, DQN, and PSO methods, respectively. The study demonstrates enhanced convergence speed and real-time response capabilities, offering a practical data-driven solution for voltage control in modern distribution networks with high renewable integration.