Prioritized sum-tree experience replay TD3 DRL-based online energy management of a residential microgrid

Abstract

Online energy management utilizing the real-time information of a residential microgrid (RM) can make full use of renewable energy and demand-side resources at the residential level. However, existing online energy management methods for RMs have poor robustness against environmental changes, which limits their applicability in highly uncertain scenarios. To address this, a novel online energy management method based on the prioritized sum-tree experience replay strategy with a double delayed deep deterministic policy gradient (PSTER-TD3) is proposed in this paper. First, we formulate the sequential scheduling decision problem as a Markov decision process (MDP) problem with the objective of minimizing residential energy costs while simultaneously ensuring household thermal comfort and minimizing range anxiety for electric vehicle usage. Then, using the proposed method, we determine the optimal online scheduling strategy under this objective. By integrating the prioritized experience replay strategy of the summation tree structure into TD3, the agent is able to learn the optimal scheduling strategy in complex environments, and its optimization performance and policy learning efficiency are significantly improved. In addition, its ability to handle multidimensional continuous action spaces helps achieve finer-grained optimization for RMs. The case study results demonstrate that the proposed method can effectively reduce the energy costs of residential microgrids while satisfying household thermal comfort requirements and reducing range anxiety for electric vehicle usage. Moreover, the optimization performance of the proposed method is robust when the uncertainty factors fluctuate violently in the environment. • A MDP with unknown state transition probabilities is established. • the thermal comfort and range anxiety are considered in the HEMS joint scheduling. • A RM online energy management method based on a novel DRL method is proposed. • Results show the superior performance of the proposed energy management method.