A Data-Driven Vehicle Speed Prediction Transfer Learning Method With Improved Adaptability Across Working Conditions for Intelligent Fuel Cell Vehicle
Dagang Lu, Donghai Hu, Jing Wang, Wenxuan Wei, Xiaoyan Zhang
Abstract
Predictive energy management strategy for intelligent fuel cell vehicle low energy consumption relies on data-driven vehicle speed prediction to provide accurate short-term speed. The traditional data-driven vehicle speed prediction based on the vehicle speed sequence lacks consideration of the influence of the power sequence of the power source on the prediction accuracy in fuel cell vehicle. Most data-driven vehicle speed prediction is based on deep learning methods, which have poor generalization capabilities when predicting across working conditions. Based on this, this paper proposes a transfer learning method for vehicle speed prediction that improves across working conditions adaptability. First, Obtain vehicle data under urban, congestion, suburban and expressway conditions through real vehicle road testing. Secondly, a vehicle speed prediction model based on hybrid deep learning is established. And the impact of dual power source power sequence on vehicle speed prediction performance is analyzed under urban conditions. Finally, a vehicle speed transfer prediction method that integrates hybrid deep learning and transfer learning is proposed. And the prediction results of vehicle speed transfer from urban conditions to congestion, suburban and expressway conditions are analyzed. The results show that compared without transfer learning. The vehicle speed transfer prediction method proposed reduces mean absolute error by 21.3%, 24.8%, and 24.9%, respectively, and mean square error by 35.3%, 44.6%, and 37.2% respectively when the prediction time domain is 5 seconds. Both mean absolute error and mean square error are reduced. This method has higher vehicle speed prediction accuracy and better adaptability across working conditions.