A Decentralized Multi-agent Energy Management Strategy Based on a Look-Ahead Reinforcement Learning Approach
- Arash Khalatbarisoltani - Université du Québec à Trois-Rivières, Department of Electrical and Computer Engineering, Canada ,
- Mohsen Kandidayeni - Université de Sherbrooke, Energy Storage and Conversion Lab (e-TESC), Canada ,
- Loic Boulon - Université du Québec à Trois-Rivières, Department of Electrical and Computer Engineering, Canada ,
- Xiaosong Hu - Chongqing University, Department of Automotive Engineering, Canada Cranfield University, Advanced Vehicle Engineering Centre, UK
Journal Article
14-11-02-0012
ISSN: 2691-3747, e-ISSN: 2691-3755
Sector:
Citation:
Khalatbarisoltani, A., Kandidayeni, M., Boulon, L., and Hu, X., "A Decentralized Multi-agent Energy Management Strategy Based on a Look-Ahead Reinforcement Learning Approach," SAE Int. J. Elec. Veh. 11(2):151-164, 2022, https://doi.org/10.4271/14-11-02-0012.
Language:
English
Abstract:
An energy management strategy (EMS) has an essential role in ameliorating the
efficiency and lifetime of the powertrain components in a hybrid fuel cell
vehicle (HFCV). The EMS of intelligent HFCVs is equipped with advanced
data-driven techniques to efficiently distribute the power flow among the power
sources, which have heterogeneous energetic characteristics. Decentralized EMSs
provide higher modularity (plug and play) and reliability compared to the
centralized data-driven strategies. Modularity is the specification that
promotes the discovery of new components in a powertrain system without the need
for reconfiguration. Hence, this article puts forward a decentralized
reinforcement learning (Dec-RL) framework for designing an EMS in a heavy-duty
HFCV. The studied powertrain is composed of two parallel fuel cell systems
(FCSs) and a battery pack. The contribution of the suggested multi-agent
approach lies in the development of a fully decentralized learning strategy
composed of several connected local modules. The performance of the proposed
approach is investigated through several simulations and experimental tests. The
results indicate the advantage of the established Dec-RL control scheme in
convergence speed and optimization criteria.