Scholarlink

<p>Charging data provide important information for monitoring and estimating the status of electric vehicle (EV) batteries. However, obtaining complete data during the charging process is difficult because of the unpredictability of user charging behavior and the diversity of charging starts and ends during the operation of EVs. There is a practical need to use short charging data segments to predict the data of the entire charging process. On the basis of the matrix long short-term memory (mLSTM), Transformer, and Mamba models, a multi-scale feature promoted expert prediction model is designed to predict the multivariate charging curve of EV battery packs. This model first extracts multi-scale seasonal and trend components from the input data, and further extracts contextual information via mLSTM. Then, a hybrid Mamba-Transformer expert module is used to capture the long-term and short-term patterns, and dynamically learn the contributions of two experts to provide prediction results. The charging data of 20 on-road EVs are used for testing. The test results show that the designed model can predict charging data for multiple parameters such as the SOC, voltage, and current, with an input length of 136 sampling points. The mean absolute error is only 0.174 for all tested EVs, indicating that the proposed model outperforms compared models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multivariate charging curve prediction of electric vehicle battery pack via multi-scale features promoted Mamba-Transformer expert

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • Yanli Yang,
  • Yizhe Jia

摘要

Charging data provide important information for monitoring and estimating the status of electric vehicle (EV) batteries. However, obtaining complete data during the charging process is difficult because of the unpredictability of user charging behavior and the diversity of charging starts and ends during the operation of EVs. There is a practical need to use short charging data segments to predict the data of the entire charging process. On the basis of the matrix long short-term memory (mLSTM), Transformer, and Mamba models, a multi-scale feature promoted expert prediction model is designed to predict the multivariate charging curve of EV battery packs. This model first extracts multi-scale seasonal and trend components from the input data, and further extracts contextual information via mLSTM. Then, a hybrid Mamba-Transformer expert module is used to capture the long-term and short-term patterns, and dynamically learn the contributions of two experts to provide prediction results. The charging data of 20 on-road EVs are used for testing. The test results show that the designed model can predict charging data for multiple parameters such as the SOC, voltage, and current, with an input length of 136 sampling points. The mean absolute error is only 0.174 for all tested EVs, indicating that the proposed model outperforms compared models.