A Dual-Evaluation Collaborative Decision Deep Reinforcement Learning Strategy for Multi-Constraint Energy Management of Hybrid Electric Vehicles

Coordinated control of fuel economy, state-of-charge (SOC), thermal state, and engine operation is central to hybrid electric vehicles (HEVs) energy management. In existing discrete-action deep reinforcement learning (DRL) methods, action selection usually relies on the Q-values predicted by the value network, whereas battery SOC and thermal constraints are often incorporated only through reward penalties after action execution. Accordingly, the short-term consequences of candidate actions cannot be compared directly during action screening, and the learned policy may exhibit SOC drift, thermal buildup, and frequent engine start-stop events. To address this issue, this paper proposes RAEP-KAN-DQN. The method retains the value-function framework and introduces executable-action mapping, one-step state prediction, and risk-aware action selection. Experimental results show that the main strength of RAEP-KAN-DQN lies in its balanced multi-objective performance. Under the standard condition, the method reduces fuel cost by 33.58% relative to Deep Q-network (DQN), while lowering the peak core and surface temperatures by 11.73% and 10.99%, respectively. When the initial temperature is raised to 40°C, fuel cost decreases by 13.61% relative to the standard-temperature case, and core and surface temperature rises are 62.13% and 62.66% lower than those of DQN. To further evaluate driving-cycle robustness, the HWFET driving cycle is adopted as a high-speed continuous-load condition. The proposed method limits SOC deviation to 0.111, which is 54.32% lower than that of DQN, while maintaining better thermal behavior and engine operation. These results show that the proposed dual-evaluation decision strategy can improve the coordinated control capability of DRL-based energy management under multiple constraints.