Power-Aware Predictive Energy Management Integrating a Transformer-Based Hybrid Prediction Method for Fuel Cell Vehicles

Predictive energy management strategies (PEMSs) offer significant potential to improve the performance of electrified vehicles by optimizing energy allocation based on anticipated driving conditions. However, conventional PEMS depend predominantly on vehicle speed prediction as a surrogate for power demand, an indirect approach that struggles to capture the complexity of real-world energy requirements. To address this limitation, this study proposes a direct multistep power demand prediction framework for PEMS, integrating a Transformer-based predictor with a physical dynamics model. The Transformer predicts future acceleration and gradient sequences from historical driving data, which are then integrated into the physical model to directly estimate multistep power demand. These predictions are sent into a deep reinforcement learning (DRL) optimizer, enabling more effective energy management. The proposed framework was validated on a fuel cell electric vehicle across diverse road conditions, optimizing energy economy, component durability, and cabin comfort. Results show that our hybrid prediction method improves power prediction accuracy by 5.3% compared to state-of-the-art methods, while the proposed PEMS achieves 3.21%–5.62% better energy efficiency than existing speed- and gradient-based PEMS. Furthermore, it outperforms single-step power prediction methods by 2.28%–5.40%, highlighting its superior adaptability. Importantly, the proposed PEMS exhibits robust performance across multiobjective tasks, confirming its potential for real-world deployment.