Reinforcement Learning Based Optimal Energy Management of A Microgrid

Modern day power systems are characterized by bi-directional power flows, e.g. vehicle-to-grid (V2G) and peer-to-peer (P2P) energy sharing systems. However, the energy scheduling is pivotal to ensure the optimal resource sharing and utilization in these interconnected systems. Renewable energy based distributed generators (DGs) have stochastic outputs and the customer typically produces excess energy comparing to the load. The imbalance between generation and load introduces complexities in the power systems as number of decision variables can be substantially increased for scheduling algorithms. There is a need of an intelligent energy management system (EMS) where we can deal with the stochastic nature of the microgrids (MGs). Recently, the reinforcement learning (RL) based techniques gain a significant interest in MG's scheduling problems. In this paper, we propose a RL based generic environment and explored two model free algorithm for scheduling problem where the RL agent can incorporate a set of constraints and energy balance rules across the system. Proximal Policy Optimization (PPO) as on-policy and Deep Deterministic Policy optimization (DDPO) as off-policy are two RL algorithms used to explore and analyze the system's performance and stability. Both the algorithms showed a great performance as both achieved the system stability with a speedy processing time.