重庆理工大学学报2025，Vol.39Issue(19)：94-101,8.DOI:10.3969/j.issn.1674-8425(z).2025.10.011
结合特征组合与MHSA-LSTM的氢燃料电池车辆单体电压最高值预测
Prediction of the highest single cell voltage in hydrogen fuel cells based on feature combination using MHSA-LSTM
摘要
Abstract
Hydrogen fuel cell vehicles have great potentials in advancing urban public transport.Compared to other new energy vehicles,hydrogen fuel cell vehicles offer advantages such as longer driving range,higher energy efficiency,and faster refueling.Thus,they have gained keen academic interest.Nevertheless,their safety risks should not be ignored.The maximum voltage of the cell unit serves as one of the indicators for evaluating the safety of power batteries in hydrogen fuel cell vehicles,yet accurately predicting this remains challenging under complex operation conditions. Although fault detection for fuel cell vehicles has seen significant progress,predicting the maximum single-cell voltage accurately remains difficult due to the complex nonlinear and dynamic coupling of electrochemical,thermal,and fluidic variables.A high-dimensional feature space,where stack current,temperature,gas pressure,and humidity interact in a strongly interdependent manner,affects the voltage output.These interactions create intricate nonlinear relationships that resist decoupling.They also cause a"short-board effect",where the weakest cell dictates the stack's overall performance and durability.Voltage uniformity,which deteriorates under high current and dynamic loading,becomes critical yet difficult to model using conventional linear methods.Moreover,the statistical distribution of cell voltages often shows skewness,and the system's dynamic response introduces local maxima in voltage variation during rapid load changes,adding complexity that traditional physical or equivalent circuit models struggle to capture. To address these limitations in existing fault detection methods for the power batteries of hydrogen fuel cell vehicles,this paper proposes a hybrid prediction model integrating long short-term memory(LSTM)network and multi-head self-attention(MHSA)mechanism.The model forecasts the peak voltage of power batteries in hydrogen fuel cell buses.Using the Spearman correlation coefficient,the method identifies and removes irrelevant and redundant features based on their correlation with the maximal cell voltage.An LSTM network extracts temporal features,while the MHSA mechanism enhances the weights of key features.A fully connected layer then employs the processed features for regression prediction.Comparative models such as GRU and LSTM are used for analysis.Simulation experiments using a hydrogen fuel cell vehicle driving dataset show MHSA-LSTM model achieves higher prediction accuracy and stability than the baseline models,offering a decision-making basis for over-voltage alarms for power batteries of hydrogen fuel cell vehicles.关键词
氢燃料电池车辆/电池单体电压/最高值预测/MHSA-LSTM组合模型Key words
hydrogen fuel cell vehicles/cell voltage/highest value prediction/MHSA-LSTM分类
信息技术与安全科学引用本文复制引用
何梓豪,孙仁云,蒋康,程培高..结合特征组合与MHSA-LSTM的氢燃料电池车辆单体电压最高值预测[J].重庆理工大学学报,2025,39(19):94-101,8.基金项目
四川省科技厅重点研发项目(23ZDYF0506) （23ZDYF0506）
成都市科技局重点研发项目(2022-YF05-01047-SN) （2022-YF05-01047-SN）
西华大学研究生科创竞赛项目(2024-227) （2024-227）
