A knowledge-assisted deep reinforcement learning approach for energy management in hybrid electric vehicles

Achieving a robust Energy Management Strategy (EMS) for Hybrid Electric Vehicles (HEVs) requires meeting several control objectives, such as drivability, reducing fuel consumption, and maintaining the state of battery charge (SoC), making the EMS a critical component of HEVs. The current EMS relies on prior knowledge of the instantaneous optimal working points of the Internal Combustion Engine (ICE), which leads to suboptimal solutions in episodic driving cycles. Previous efforts to implement Reinforcement Learning (RL) encountered challenges such as slow convergence, instability in tracking driving cycles, and inadequate performance under real driving conditions. This paper presents an intelligent EMS based on a hybrid Knowledge-Assisted system that integrates a Deterministic Policy Gradient (KA-DDPG) and a Deep Q-Network (KA-DQN) to overcome the challenges of RL and achieve optimal EMS actions under various driving conditions. Two versions of the proposed algorithm—offline and online—are presented. Simulation results show that KA-DDPG requires less computation time, reduces fuel consumption by 6.99 %–7.26 % in offline mode and 5.18 %–5.67 % in online mode, and maintains SoC stability. These methods improve average negative electric motor torque and result in greater energy savings, while the robustness of the algorithm has been examined by changing the vehicle's weight.