Benchmarking Deep Reinforcement Learning Based Energy Management Systems for Hybrid Electric Vehicles

Abstract

Energy management strategy (EMS) is important for improving the fuel economy of hybrid electric vehicles (HEVs). Deep reinforcement learning techniques have seen a great surge of interest, with promising methods developed for hybrid electric vehicles EMS. As the field grows, it becomes critical to identify key architectures and validate new ideas that generalize to new vehicle types and more complex EMS tasks. Unfortunately, reproducing results for state-of-the-art deep reinforcement learning-based EMS is not an easy task. Without standard benchmarks and tighter metrics of experimental reporting, it is difficult to determine whether improvements are meaningful. This paper conducts an in-depth comparison between numerous deep reinforcement learning algorithms on EMSs. Two different types of hybrid electric vehicles, which include an HEV with planetary gears for power split and a plug-in HEV, are considered in this paper. The main criteria for performance comparison are the fuel consumption, the state of batteries’ charges, and the overall system efficiency. Moreover, the robustness, generality, and modeling difficulty, which are critical for machine learning-based models, are thoroughly evaluated and compared using elaborate devised experiments. Finally, we summarize the state-of-the-art learning-based EMSs from various perspectives and highlight problems that remain open.