For some HEVs, depending on the system configuration, the integration degree of vehicle control strategies, and modeling methods, both discrete and continuous actions can exist in the same action space, making it difficult to describe them monolithically by either discrete action space or continuous action space. Taking a power-split hybrid electric bus (HEB) as an example, this chapter will introduce how to address EMS learning problems in such hybrid action spaces by combing the idea of action value learning and policy gradient update. Furthermore, an energy management method considering terrain information is described, and accordingly, the influence of the multi-source information on learning-based EMSs is discussed in terms of fuel economy, strategy performance under specific driving scenarios, and the strategy decisions.
Learning of EMSs in Discrete-Continuous Hybrid Action Space
Deep Reinforcement Learning-Based Energy Management for Hybrid Electric Vehicles ; Kapitel : 5 ; 77-99
01.01.2022
23 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
ELBA - Bundesanstalt für Straßenwesen (BASt) | 2022