Solving large-scale combined heat and power economic dispatch problems by using deep reinforcement learning based crisscross optimization algorithm

Abstract

The extensive interconnection in bulk power systems brings great challenges in the economic dispatch problems, especially the large-scale combined heat and power economic dispatch (CHPED), which is quite intractable due to complex thermal and electrical couplings in cogeneration units. Most meta-heuristic algorithms cannot work well while considering non-convexity, discontinuity, and non-differentiability. Particularly, when it comes to the high dimensional and large-scale CHPED problems, these algorithms are either easy to trap in the local optimum or quite time-consuming. To address these issues, a novel deep reinforcement learning (DRL) based crisscross optimization (CSO) algorithm is proposed for the first time, which can (1) improve the searching ability of CSO, (2) enhance the overall searching efficiency, and (3) perform extraordinarily well for the large-scale CHPED problems. Firstly, to further improve the global search ability, DRL is employed to reduce dimensionality and narrow the search space of the initial population of the CSO algorithm. Then, DRL is updated by the deep deterministic policy gradient (DDPG) strategy, which can facilitate an instantaneous decision in accelerating CSO. Thereafter, by adopting each decision provided by DRL, a great difference may occur in the same scenario to speed up the crossover procedure and guide the upcoming search round. On this basis, the enhanced CSO algorithm could quickly carry out crossovers with search space contraction and dimensionality reduction. Finally, experiments on four test systems have substantiated the effectiveness of the proposed algorithm, especially the superiority over other state-of-the-art techniques in solving large-scale CHPED problems.