Bridging Representation Learning and Optimal Control: Joint Representation–Decision Optimization for FCEV Energy Management

It is commonly assumed that future driving condition prediction is beneficial for energy management strategies (EMSs) for fuel cell electric vehicles (FCEVs). Yet, this intuition does not always hold: improving prediction accuracy does not necessarily yield better control performance, especially for deep reinforcement learning (DRL)-based downstream policies. Motivated by this observation, this paper proposes a control-oriented framework based on joint representation-decision optimization. A temporal convolutional network (TCN) is used to encode historical driving data into a context representation and an auxiliary prediction output, where the context representation augments the state for DRL-based decision making. During training, the representation is jointly shaped by control feedback and auxiliary prediction regularization, so that temporal consistency is preserved while decision-relevant information is enhanced. Simulation results under the WLTC cycle show that the proposed method reduces the total operating cost by 9.51% compared to a TCN-based model predictive control (MPC) benchmark and outperforms state-of-the-art predictive DRL baselines. Pearson correlation analysis further shows that the learned context representations are more strongly associated with control objectives than prediction-oriented representations, indicating that the proposed framework drives context features beyond predictive information toward more decision-relevant representations.