Enhancing Energy Management Strategy for Battery Electric Vehicles: Incorporating Cell Balancing and Multi-Agent Twin Delayed Deep Deterministic Policy Gradient Architecture
Armin Lotfy, Hicham Chaoui, Mohsen Kandidayeni, Loïc Boulon
Abstract
This paper introduces a real-time multi-objective adaptive Energy Management Strategy (EMS) based on a Multi-Agent Reinforcement Learning (MARL) architecture. Leveraging Twin Delayed Deep Deterministic Policy Gradient (TD3) methods, this EMS continuously monitors the system, striking a balance between front and rear electric drive operations, cell balancing in batteries, and other crucial parameters affecting battery aging. It not only meets driver requirements but also determines the optimal power levels for Electric Motors (EMs), reducing battery depletion and aging. Validation employs a 2021 Motor Vehicle Challenge model with two electric motors. Results indicate the advantages of the proposed EMS, meeting driver power needs across diverse environmental conditions. Furthermore, it achieves a final state of charge (SOC) within a mere 0.3% deviation from the Dynamic Programming (DP) approach. The EMS excels by effectively balancing battery cells and optimizing temperature, mitigating long-term battery aging. Importantly, it outperforms the highest reported SOC value in the 2021 Motor Vehicle Challenge while satisfying all specified criteria.