Superior energy management for fuel cell vehicles guided by improved DDPG algorithm : integrating driving intention speed prediction and health-aware control

Abstract

Despite the significant advantages of fuel cell (FC) vehicles in reducing urban air pollution and extending driving range, effectively managing their internal energy systems remains a major challenge. To maximize the operational efficiency and lifespan of the FC system without compromising fuel economy, this paper proposes a novel predictive energy management paradigm guided by deep reinforcement learning. This strategy innovatively integrates driving intention speed prediction and health-aware control. Specifically, we developed a multi-input bi-directional long short-term memory (BiLSTM) predictor incorporating driving intentions (DI-BiLSTM) using the fuzzy C-means algorithm to enhance the prediction accuracy of future vehicle state trajectories. Downstream control decisions are executed through an improved deep deterministic policy gradient (DDPG) algorithm, which optimizes action space selection based on the degradation characteristics of the FC system. Additionally, during the training and validation phases of the energy management strategy (EMS), we utilized high-quality driving data collected from real bus routes using a high-performance Beidou integrated navigation system, replacing conventional standard driving cycles to enhance the strategy's generalization ability across different scenarios. The results indicate that, compared with conventional prediction model relying solely on historical speed data, the DI-BiLSTM improves prediction accuracy by at least 7.86 % over 3 s, 5 s, and 8 s prediction horizons. Compared with conventional DDPG-based EMS, the proposed EMS increases the average efficiency of the FC system by 32.18 % and extends its lifespan by 16.50 %. In terms of overall driving costs, the proposed EMS improves driving economy by 9.97 % compared with conventional DDPG-based EMS.