A Deep Reinforcement Learning Framework for Adaptive Resiliency Enhancement in Smart Power Grids

This paper addresses the critical challenge of enhancing resiliency in smart power grids against cyber-physical disturbances, including cyber-attacks, system faults, and dynamic load variations. Traditional control methods often lack adaptability and robustness required to maintain stable grid operation under such evolving threats. To overcome these limitations, we propose a deep reinforcement learning framework based on the Deep Deterministic Policy Gradient (DDPG) algorithm, optimized using the Root Mean Square Propagation (RMSprop) method for stable and efficient training. The proposed framework features a two-layered control architecture that models grid states, actions, and rewards to enable continuous, real-time adaptive control of distributed energy resources, microgrids, and control centers. Key contributions include the integration of DDPG with RMSprop optimizer tailored for power system stability, a comprehensive reward design reflecting resiliency metrics, and a resiliency adaptation layer for rapid response to disturbances and cyber-attacks. Simulation results conducted on a realistic smart grid testbed demonstrate superior performance in system stability, control accuracy, convergence speed, and resiliency score compared to classical Proportional-Integral-Derivative (PID) controllers and alternative reinforcement learning approaches such as Deep Q-Network (DQN) and Proximal Policy Optimization (PPO). The findings highlight the framework’s potential for practical deployment, offering a scalable and intelligent solution to enhance smart grid resiliency in the face of increasing cyber-physical challenges.