Simulation-Based Electrothermal Feature Extraction and FCN–GBM Hybrid Model for Lithium-ion Battery Temperature Prediction
Accurate temperature prediction plays a vital role in the thermal management and safety assurance of lithium-ion battery systems. This study proposes a hybrid temperature prediction model that integrates Fully Connected Networks (FCN) and Gradient Boosting Machines (GBM) to capture temperature evolu...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
The Electrochemical Society of Japan
2025-08-01
|
| Series: | Electrochemistry |
| Subjects: | |
| Online Access: | https://www.jstage.jst.go.jp/article/electrochemistry/93/8/93_25-00067/_html/-char/en |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Accurate temperature prediction plays a vital role in the thermal management and safety assurance of lithium-ion battery systems. This study proposes a hybrid temperature prediction model that integrates Fully Connected Networks (FCN) and Gradient Boosting Machines (GBM) to capture temperature evolution under varying discharge rates. A nonlinear electrothermal simulation framework based on the Nonlinear Thermal Generalized Kalman (NTGKF) model is first constructed to analyze the electrothermal coupling behavior of batteries under discharge rates ranging from 1 to 5 times the nominal capacity. Leveraging the nonlinear feature extraction capability of FCN and the ensemble learning robustness of GBM, an FCN–GBM hybrid model is developed and evaluated using different input configurations, including voltage alone, internal resistance alone, and the combination of both. Simulation and prediction results demonstrate that the evolution of voltage and internal resistance closely aligns with temperature variation, indicating their suitability as key features for temperature modeling. The proposed FCN–GBM model is applied to predict discharge temperature profiles of LiFePO4 (LFP) and LiNi0.8Co0.15Al0.05O2 (NCA) cells. The combination of voltage and resistance as input features significantly enhances prediction performance. Under the 20 % training data condition, the LFP model achieves a mean absolute error (MAE) of 0.4576 K and a root mean square error (RMSE) of 0.5411 K, compared to 1.3404 K and 1.5727 K for the baseline GBM model. For the NCA battery, the FCN–GBM model achieves an MAE of 0.3025 K and an RMSE of 0.9973 K, also outperforming GBM with respective errors of 0.8447 K and 1.5878 K. Moreover, the use of single input features leads to larger prediction errors. These results confirm the effectiveness of the FCN-enhanced GBM model and the advantage of feature fusion using voltage and resistance, providing practical insights for improving thermal management and risk mitigation in lithium-ion battery systems. |
|---|---|
| ISSN: | 2186-2451 |