Sensitivity Analysis of Reinforcement Learning-Based Hybrid Electric Vehicle Powertrain Control

Hybrid Electric Vehicles (HEVs) achieve better fuel economy than conventional vehicles by utilizing two different power sources: an internal combustion engine and an electrical motor. The power distribution between these two components must be controlled using some algorithm, be it rule based, optimization based, or reinforcement learning based. In the design of such control algorithms, it is important to evaluate the impact that variations of certain design parameters will have on the system performance, in this case, fuel economy. Traditional methods of sensitivity analysis have been applied to various power flow control algorithms to determine their robustness to the variations of HEV design parameters. This article presents a sensitivity analysis of three power flow control algorithms: twin delayed deep deterministic policy gradient (TD3), deep deterministic policy gradient (DDPG), and adaptive equivalent consumption minimization strategy (A-ECMS). The overall results show that the deep reinforcement learning (DRL)-based control algorithms have similar robustness, but higher design predictability compared to the conventional A-ECMS algorithm.