Deep reinforcement learning applied to Monte Carlo power system reliability analysis

This paper presents a proof of concept for a novel three-step method for using reinforcement learning in Monte Carlo simulations for power system reliability studies. Traditionally, an optimal power flow solver is used to compute the remedial actions necessary after a power system failure. Our new approach provides these remedial actions by the policy of a deep reinforcement learning agent. To train this agent, we construct a new power reliability reinforcement learning environment. After training the agent using the Twin Delayed Deep Deterministic Policy Gradient algorithm with a prioritized experience replay buffer, the agent is used within a time sequential composite power system Monte Carlo reliability simulation. We show that when the trained agent is used to find the remedial measures in these Monte Carlo simulations, we obtain near-optimal results on two standard power reliability test cases with favorable scaling of computation time. This work is a first step in using reinforcement learning in reliability analysis and establishes a foundation for future work.