Deep Reinforcement Learning for Real-Time Energy Management and Optimal Power Allocation in Distributed Electrical Systems

Abstract

The increasing penetration of renewable energy resources and distributed generation has made real-time energy management in microgrids more complex due to fluctuating demand, intermittent generation, dynamic electricity prices, and energy-storage constraints. The proposed approach considers the operation of a microgrid as a Markov Decision Process with states: electrical demand, renewable generation, electricity price, energy-storage state of charge, and temporal operating condition. Continuous control actions for battery charging and discharging, dispatchable generation and grid power exchange are generated using a Soft Actor-Critic algorithm. The goal of the control is to reduce operating cost, renewable energy curtailment, power imbalance, network losses, and constraint violations, while ensuring reliable system operation. Highly variable operating conditions are accounted for, such as varying PV and wind generation, variable load demand, changing electricity prices, periods of renewable surplus and varying storage bounds. The proposed framework is tested on an independent training and out-of-sample testing set, to evaluate the ability of the framework to generalize in unseen operating conditions. Conditions are evaluated based on economic, technical, renewable, storage, and computational metrics such as operating cost, grid reliance, renewable energy usage, power losses, power state-of-charge, constraint violations, and decision time. The framework serves as a baseline to compare adaptive DRL-based energy management with traditional and other reinforcement-learning approaches and serves as a replicable foundation for assessing the approach.