Battery Life Evaluation Method Based on Temporal Convolution Network
ObjectiveTo improve the technical economy of battery system applications, a temporal convolutional network (TCN) is employed to evaluate battery life from two perspectives: State of health (SOH) and remaining useful life (RUL). SOH is typically quantified by capacity, while RUL is measured in terms...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Editorial Department of Journal of Sichuan University (Engineering Science Edition)
2025-07-01
|
| Series: | 工程科学与技术 |
| Subjects: | |
| Online Access: | http://jsuese.scu.edu.cn/thesisDetails#10.12454/j.jsuese.202300930 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | ObjectiveTo improve the technical economy of battery system applications, a temporal convolutional network (TCN) is employed to evaluate battery life from two perspectives: State of health (SOH) and remaining useful life (RUL). SOH is typically quantified by capacity, while RUL is measured in terms of the remaining number of charge-discharge cycles.MethodsFirst, the TCN-based approach to battery life assessment is introduced. Compared to classical recurrent neural networks, TCN has the advantages of improved gradient stability, faster data processing, and reduced memory consumption. Next, 14 indirect health-related features are extracted from readily available battery data including time, voltage, current, and temperature. The relevance of each feature to capacity is assessed using three correlation techniques: KL divergence, Pearson correlation coefficient, and gray relational analysis. Additionally, a polling strategy is employed using the TCN: each feature is used individually for capacity prediction, and the mean of six prediction outcomes per feature is taken as the final correlation score. Correlation results from all four methods are compared. Although Pearson correlation and TCN-based analysis yield similar rankings, both focus primarily on the top features. Due to the different principles of these methods, their outcomes often diverge significantly. Therefore, TCN-derived results are considered the most reliable for identifying factors influencing capacity prediction. For lithium-ion battery data from NASA, five features with the greatest impact on SOH are identified (in descending order of importance): cycle time, average voltage, voltage sample entropy, temperature sample entropy, and current.Results and DiscussionsTo address redundancy caused by overlapping information among these features, Kernel principal component analysis (KPCA) is applied for dimensionality reduction. The contribution rates of principal components are calculated, and the top two principal components (PC1 and PC2) are selected as inputs for simulation to eliminate noise and improve computational efficiency. A comparative analysis of three prediction models including TCN, long short-term memory (LSTM) networks, and Backpropagation (BP) neural networks shows that the TCN achieves the lowest root mean square error (RMSE) of 0.019 3, indicating the highest predictive accuracy. Battery capacity regeneration can result in similar capacities at different numbers of cycles or times. When SOH is used to characterize battery life, significant errors will be introduced. In contrast, RUL will irreversibly decrease as the service time increases, which can provide a more reliable criterion for evaluation. To predict RUL, the same five key features (cycle time, average voltage, voltage sample entropy, temperature sample entropy, current) along with capacity are used. Six principal components are obtained through KPCA analysis, and PC1 and PC2 are also selected for prediction. Among the three models compared, the RMSE of TCN is lowest (0.019 3), which further proves its outstanding accuracy in RUL estimation.ConclusionsThis study validates the effectiveness of TCN for battery life prediction. By evaluating both SOH and RUL, the proposed TCN-based framework provides a more accurate and robust assessment of battery health. |
|---|---|
| ISSN: | 2096-3246 |