A constrained DRL-based bi-level coordinated method for large-scale EVs charging. (1st February 2023)

• Record Type: Journal Article
• Authors: Ming, Fangzhu; Gao, Feng; Liu, Kun; Li, Xingqi
• Is Part Of: Applied energy. Volume 331(2023)
• Publication Date: 2023-02-01
• Subjects: Stochastic scheduling -- Multi-energy system -- Hierarchical optimization -- Constrained Markov decision processes -- Deep reinforcement learning; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2022.120381
• Languages: English
• ISSNs: 0306-2619

Abstract

Abstract: With the vigorous development of battery electric vehicles (BEVs), BEVs' charging scheduling is essential for better economy and safety. In this paper, we aim to minimize the electricity purchasing cost considering a large number of BEVs and distributed energy. This problem is challenging to get the optimal charging policy due to a large number of uncertainties and dimension disasters caused by a large scale of BEVs and renewable energy. To meet these challenges, we propose an improved bi-level schedule framework, which decomposes the primal problem into two sub-problems to reduce the computational complexity and designs a communication mechanism to ensure the consistency of optimality between different levels. Then the problem is modeled as constrained multi-level Markov decision processes (CMMDP). In the upper level, a constrained deep reinforcement learning method (CDRL) is proposed to get the total charging or discharging energy of BEV groups. An action constraint module is constructed to ensure the feasibility of chosen actions and a novel reward shaping function is designed to optimize action selection. In the lower level, an optimal descending order charging policy (DOCP) is taken to fast decide the charging or discharging behavior for each BEV based on the upper level's decision. Numerical experiments show that our method has obvious superiority in training efficiency and solution accuracy compared with state of art DRL methods, and reduces the cost by 12% to 28% compared with an experience charging policy. Highlights: A new bi-level scheduling framework is proposed for a large-scale charging problem. A CMMDP model is built to describe the stochastic multi-stage charging problem. A Constrained DRL method considering action constraints is proposed in upper level. A heuristic charging policy is taken to solve the dimension curse in lower level. A communication module ensures the feasibility and consistency of solutions.