摘要
通过对锂离子电池容量退化轨迹的精确预测可以大幅提升电池材料的研究效率。针对Transformer网络在锂电池容量退化轨迹预测这种长时间序列预测任务中存在的问题,本工作采用滑动窗口策略,构建了一种基于Informer网络的锂离子电池容量退化轨迹预测方法。首先,利用滑动窗口对数据集进行划分和再拼接,便于神经网络挖掘数据序列内部的相关性;然后,根据Informer网络的周期性时间特征捕捉能力设计适用于锂电池数据的全局时间戳;最后,使用前10%容量数据通过多步滚动预测方法实现模型输出,缓解预测中的误差累积问题,进而得到完整的预测轨迹。通过选取不同的误差评价指标和训练过程中的时间开销,在美国马里兰大学提供的锂电池数据集上验证了所建立模型的准确性和训练效率,并在美国航空航天局提供的电池数据集上验证了模型的泛用性。本工作模型的预测结果与多层感知机神经网络、循环神经网络及Transformer网络模型对比,退化轨迹与真实轨迹最为拟合,且训练时间开销小,预测结果的平均绝对误差和均方根误差控制在2.57%和3.5%,验证了所提预测方法的有效性。
Accurate prediction of lithium-ion battery capacity degradation trajectories enhances the efficiency of battery materials research.Aiming to resolve the challenges associated with the Transformer network in the prediction of lithium-ion battery capacity degradation trajectory,this study adopts the sliding window strategy and constructs a lithiumion battery capacity degradation trajectory prediction method based on Informer,a time series forecasting model.First,the sliding window is used to divide and re-splice the dataset;this facilitates the neural network to exploit the correlation within the dataset;subsequently,the global timestamp applicable to lithium-ion battery data is designed according to the periodic time series capturing ability of Informer;finally,the model output is realized through the multistep rolling prediction method by using the first 10%of the battery capacity data to alleviate the error accumulation in the prediction,subsequently obtaining the complete prediction trajectory.The accuracy and training efficiency of the established model are verified using the lithium-ion battery dataset provided by the University of Maryland.Different error evaluation and time overhead metrics are selected in the training process;additionally,the generalizability of the model is verified using the lithium-ion battery dataset provided by NASA.Comparing the prediction results of the model in this study with that of the multilayer perceptron neural network,recurrent neural network,and Transformer network model,the following is observed:the degraded trajectories obtained in this study are best fitted to the real trajectories;the training time overhead is small;and,the average absolute and root mean square errors of the prediction results are controlled at 2.57%and 3.5%,thus verifying the validity of the proposed prediction method.
作者
唐梓巍
师玉璞
张雨禅
周奕博
杜慧玲
TANG Ziwei;SHI Yupu;ZHANG Yuchan;ZHOU Yibo;DU Huiling(School of Materials Science and Engineering,Xi'an University of Science and Technology,Xi'an 710054,Shaanxi,China)
出处
《储能科学与技术》
CAS
CSCD
北大核心
2024年第5期1658-1666,共9页
Energy Storage Science and Technology
基金
陕西省科技厅陕煤联合基金重大专项(2021JLM-28)。
