Data-driven approach for short-term power demand prediction of fuel cell hybrid vehicles

Abstract

Due to slow internal mass transport, the fuel cell is a typical time-delay control object in vehicular hybrid powertrain. To yield better control effect, the predictive control is considered as an effective solution, in which the short-term power demand of vehicle is a key input variable and must be predicted accurately. However, a time-phase mismatch phenomenon usually occurs in prediction results when using non-iterative direct prediction method, resulting in poor prediction accuracy. This study systematically explains the mechanism of the studied time-phase mismatch and proposes a novel iterative learning framework (ILF) to reduce it. Several machine learning algorithms are compared to select a proper learning core for ILF. The results show that prediction RMSE reduces up to 76.8% and 65.0% for the power and power change rate predictions, respectively, comparing with non-iterative prediction manner. The least-squares support vector machine as the learning core of ILF achieves the best performance within the shortest runtime. Moreover, the proposed ILF predictor has a good adaptability to various driving conditions through more validations. The proposed ILF has better predictable ability for the future data comparing with classical recurrent time-series prediction method. The proposed ILF is expected to improve the accuracy of vehicle load-status perception.