Adaptive distributed stochastic deep reinforcement learning control for voltage and frequency restoration in islanded...

This paper proposes an adaptive secondary control strategy for islanded AC microgrids (MGs) using Distributed Stochastic Deep Reinforcement Learning (DSDRL), targeting reliable frequency and voltage restoration alongside proportional power sharing. Unlike existing distributed stochastic control methods that often fail under communication noise and variable delays, our approach dynamically adapts to these uncertainties by integrating a control Lyapunov function (CLF)-based nominal model with a deep deterministic policy gradient (DDPG) algorithm. Distinct from prior DRL applications, our method rigorously couples Lyapunov-based stability guarantees with DRL-based gain tuning, ensuring formal stability despite stochastic disturbances and model uncertainties. This hybrid framework facilitates real-time optimization of control policies, contributing to improved resilience and robustness under realistic network conditions. Furthermore, the controller's design is independent of precise distributed generator (DG) parameters, supporting heterogeneous and uncertain systems. MATLAB/SimPowerSystems simulations demonstrate the proposed method's improved performance compared to existing techniques across a range of scenarios involving communication noise, time delays, and load variations. The proposed framework offers a theoretically grounded, scalable, and adaptable solution for secondary control of MGs, paving the way for future extensions to larger systems and secure control applications.