Real-time multi-objective energy management for fuel cell vehicles via temporal-frequency fusion speed prediction and local optimal SOC feedback

To enhance the battery state of charge (SOC) maintaining capability, fuel economy, and dynamic stability of fuel cell vehicles (FCVs) under different driving conditions and varying SOC states, this study proposes an adaptive equivalent consumption minimization strategy with local optimal SOC feedback (LOA-ECMS) based on Temporal-Frequency fusion speed prediction. A Fourier transform is first used to analyze the frequency spectrum of the speed profiles, validating the feasibility of frequency-domain modeling. To fully leverage the inherent periodic features within the speed profiles, a Convolutional Neural Network (CNN) based Temporal-Frequency joint feature extraction network for vehicle speed prediction (TF-VSPNet) is developed. TF-VSPNet outperforms existing mainstream speed prediction models, achieving improvements of 17.65 %, 3.59 %, and 21.05 % in MAE, RMSE, and MAPE, respectively. Based on TF-VSPNet outputs, LOA-ECMS employs an improved efficient PMP (IE-PMP) to solve the optimal SOC trajectory and adaptively adjusts the equivalent factor for predictive fuel cell power control. Simulation and experimental validations confirm that LOA-ECMS achieves superior overall control performance compared to rule-based (RB) strategy and adaptive ECMS (A-ECMS), maintaining the final SOC deviation within ±0.25 %, reducing hydrogen consumption by over 4.3 %, and suppressing fuel cell power fluctuations by more than 30 % with a peak change rate reduction above 80 %.