Data redundancy reduction using sensitivity analysis method for machine-learning-based battery management system
This thesis proposes a sensitivity analysis method to reduce the computational effort of machine-learning (ML) techniques in the battery management system (BMS). The novel approach analyzes the sensitivity of lithium-ion battery model parameters towards their discharge performances. The sensitivity...
Saved in:
| 主要作者: | |
|---|---|
| 格式: | Thesis |
| 语言: | English English English |
| 出版: |
2021
|
| 主题: | |
| 在线阅读: | http://eprints.uthm.edu.my/987/1/24p%20MUHAMMAD%20SYAFIQ%20ANWAR%20BIN%20MUSTAZA.pdf http://eprints.uthm.edu.my/987/4/MUHAMMAD%20SYAFIQ%20ANWAR%20BIN%20MUSTAZA%20COPYRIGHT%20DECLARATION.pdf http://eprints.uthm.edu.my/987/5/MUHAMMAD%20SYAFIQ%20ANWAR%20BIN%20MUSTAZA%20WATERMARK.pdf |
| 标签: |
添加标签
没有标签, 成为第一个标记此记录!
|
| 总结: | This thesis proposes a sensitivity analysis method to reduce the computational effort of machine-learning (ML) techniques in the battery management system (BMS). The novel approach analyzes the sensitivity of lithium-ion battery model parameters towards their discharge performances. The sensitivity analysis is based on the sum-of-difference method to identify redundant model parameters that characterized the battery’s discharge performance. From the sensitivity analysis, it is found out that the current, discharge time from state-of-charge (SOC), and power discharge output show minimum influence towards the variation of battery parameters. Thus, this finding indicates that current, discharge time, and power were redundant and may be excluded in the formation of the training dataset. The newly discovered finding is applied to the ML-based BMS. In the development of the training dataset, a reduced-sized dataset is formed by excluding current, discharge time, and power from the training dataset for the real-time battery-parameter monitoring in BMS. The newly formed reduced-sized dataset is applied to ML techniques: artificial neural network (ANN), deep learning (DL), and modified adaptive neuro-fuzzy inference system (MANFIS). Consequently, the training performances of all three ML techniques are observed, analyzed, and compared. The results demonstrate that the reduced-sized training dataset that is formed based on the sum-of-different method reduced the training time by up to 60.25% as compared with the full-sized dataset. Also, estimation accuracy is improved due to the improvement in training data bias. This result suggests that the proposed method significantly improved the training performance of the ML techniques in BMS application. The implementation of sensitivity analysis in the development of the training dataset for ML applications improved the performance of the real-time monitoring of lithium-ion battery parameters in advanced BMS applications. |
|---|