Testbed implementation of reinforcement learning-based demand response energy management system

Demand response (DR) has been acknowledged as an effective method to improve the stability, and financial efficiency of power grids. During operation of a DR program, there are usually multiple interactions among different grid entities, which complicates decision-making processes with respect to grid operations. Recently, reinforcement learning (RL) has attracted increasing attention for managing complex decision-making problems, owing to its self-learning capacity. Several theoretical RL-based approaches have been proposed for addressing various DR issues, but the practical feasibility of these theoretical approaches remains to be proven. In this paper, a conceptual architecture is firstly proposed to support DR management of a diversified facility in the context of a price-based DR environment. Secondly, exhaustive guidelines are provided to illustrate how to implement a multi-agent RL-based algorithm in the constructed DR management system. Afterwards, a laboratory-level testbed was set up to evaluate the effectiveness of the deployed DR algorithm. The experimental evaluation results show that the RL-based DR algorithm takes about 20s and 50 episodes to achieve optimal load control policy. By executing the optimal operation policy, the overall energy consumption during the highest price period (i.e., 15:00–18:00) is significantly reduced by 133.6% compared with the lowest price period (i.e., 02:00–05:00).