Supervised RL Boosts Distributed Energy Resource Coordination

Researchers have introduced a Supervised Reinforcement Learning (SRL) framework designed to enhance the coordination of Distributed Energy Resources (DERs). The increasing integration of DERs is vital for decarbonizing power systems, but their inherent uncertainties and complex modeling pose significant challenges for traditional optimization methods. While standard reinforcement learning (RL) offers a promising alternative, it often suffers from sample inefficiency and sub-optimality when trained from scratch. Inspired by large language model training paradigms, the proposed SRL framework addresses these issues by first pre-training a policy using demonstration data in a supervised learning manner. This initial policy is then further refined through a two-step fine-tuning process: an offline phase to boost overall performance, followed by an online phase to adapt to real-world dynamics. Experimental results demonstrate that RL implementations based on this SRL framework achieve significantly higher cost efficiency compared to all benchmarks, even when the initial demonstration data is of low quality.