This study addresses the energy management problem of fuel cell hybrid electric vehicles operating over a 562 km real-world journey across Tunisia, encompassing heterogeneous driving conditions including urban centers, highways, and rest segments. The proposed framework preserves realistic battery dynamics through continuous state-of-charge tracking, where the final state of charge of each segment initializes the next one, with natural partial recovery during rest periods. A coupled architecture combining deep reinforcement learning and a nonlinear autoregressive neural network with exogenous inputs is used to forecast critical system variables and optimize real-time power split among the fuel cell, battery, and supercapacitor. Three reward formulations are investigated and compared: a baseline policy, a battery-conservative policy, and an adaptive road-aware policy. The results show that the adaptive road-aware strategy provides the best overall balance between hydrogen economy and charge sustainability. On the long drive cycle, the adaptive road-aware policy achieves the lowest hydrogen consumption, reducing hydrogen use by 8% relative to the deep deterministic policy gradient road-aware strategy and by 27.3% relative to the equivalent consumption minimization strategy, although its final state of charge is slightly lower than those two methods. On the combined generalization driving cycle, the adaptive road-aware policy reduces hydrogen consumption by 6.64% relative to the deep deterministic policy gradient baseline, while also outperforming the equivalent consumption minimization strategy by 3.4% in final state of charge and 26.5% in hydrogen consumption, confirming strong generalization to unseen driving conditions, which demonstrates that the proposed prediction-informed reinforcement learning framework is effective for both long-distance operation and broader real-world deployment.
Sayah et al. (Mon,) studied this question.