基于极端梯度提升的PEMFC长短期老化趋势预测
SHORT-AND LONG-TERM DEGRADATION PREDICTION FOR PROTON EXCHANGE MEMBRANE FUEL CELL BASED ON EXTREME GRADIENT BOOSTING
王艳琴 1谢卓峰 2韩国鹏 1张杲 3郭爱2
作者信息
- 1. 中车唐山机车车辆有限公司,唐山 064000
- 2. 西南交通大学电气工程学院,成都 610031
- 3. 西南交通大学唐山研究院,唐山 063000
- 折叠
摘要
为了同时实现准确的燃料电池长短期老化趋势预测,提出基于极端梯度提升(XGBoost)的PEMFC老化趋势预测模型.首先,对燃料电池老化实验数据进行降噪预处理,利用双指数对电压恢复特性进行建模;然后,基于XGBoost算法,构建4种提前多步短期老化预测模型以及考虑恢复性的长期预测策略,并利用粒子群算法优化模型的参数;最后,比较4种短期预测模型的预测结果,并将最优的预测模型应用于长期老化预测策略.典型数据实验表明:采用多输入多输出策略(MIMO)的XGBoost预测模型具有最好的预测性能,其提前3步预测的均方根误差为0.00465、平均相对误差为0.00219平均运算时间为3.48 s;基于MIMO-XGBoost且考虑恢复性的长期预测策略剩余使用寿命(RUL)的平均相对误差为7.74%,显著优于自回归差分移动平均方法.
Abstract
In order to achieve accurate short-and long-term degradation prediction of fuel cells,a PEMFC degradation prediction model based on extreme gradient boosting(XGBoost)model was proposed.Firstly,the experimental data of fuel cell aging were processed to reduce noise and the voltage recovery characteristics were modeled by using double exponent.After,four multi-step ahead prediction model based on XGBoost and the long-term prediction strategy considering recoverability were constructed,and particle swarm optimization(PSO)algorithm was used to optimize the parameters of the model.Lastly,the prediction results of the four short-term prediction models were compared,and the optimal model was applied to the long-term aging prediction strategy.The results show that the XGBoost prediction model with multiple input multiple output(MIMO)strategy had the best prediction performance,which three-step ahead prediction's root mean square error was 0.00465、mean absolute error was 0.00219 and operation time was 3.48 s.The average relative error of the remaining useful life(RUL)of the long-term prediction strategy based on MIMO-XGBoost and considering recovery was 7.74%,which was significantly better than the autoregressive integrated moving average method.
关键词
燃料电池/老化/预测/剩余使用寿命/极端梯度提升Key words
fuel cells/degradation/prediction/remaining useful life/extreme gradient boosting引用本文复制引用
基金项目
国家自然科学基金(52077180)
中车十四五重大专项(2021CXZ021-4)
出版年
2024
NETL
NSTL
万方数据