Energy management strategies for fuel cell vehicles: Advances, challenges, and future directions for the prediction-control hierarchical architecture

Existing reviews of energy management strategies for fuel cell vehicles are mainly based on rule-based, optimization-based, and learning-based control approaches, with a strong focus on control-layer methodologies. There is a widespread lack of consideration for the impact of integrating vehicle speed prediction into the energy management control layer. Based on this problem, it conducts a systematic review of energy management strategies centered on vehicle speed prediction, following a prediction-control hierarchical architecture. Three evolutionary development paths are proposed and summarized: Probabilistic Prediction-Optimization Control, Intelligent Prediction-Optimization Control and Intelligent Prediction-Intelligent Control. This paper summarizes advances in the prediction layer, focusing on the fusion of multi-source spatiotemporal traffic flow information and algorithmic improvements. It also highlights breakthroughs in the control layer, such as state of charge trajectory planning, power change rate suppression, and coordinated health constraints. Based on this, three common limitations across existing studies can be summarized: At the prediction layer, speed prediction errors are amplified in a cascading manner along the chain of “speed prediction, power planning, and energy scheduling.” At the control layer, generalization for new conditions and environmental adaptability are insufficient. Additionally, the integration between prediction and control is limited, with unidirectional information transfer and misaligned objectives. This paper proposes adding an attention mechanism to the prediction layer to improve vehicle speed forecasting accuracy. It also suggests incorporating transfer learning into the control layer's strategy to enhance generalization across different operating conditions. Furthermore, end-to-end joint training enables integrated closed-loop optimization by coupling the prediction and control layers.