To address this challenge, we propose a graph-based spatiotemporal feature extraction method using multi-attention mechanism. Graph neural networks (GNNs) are particularly powerful for modeling complex data structures where relationships between nodes (features) are crucial [5]. By capturing both local and global dependencies through the graph structure, GNNs enable more accurate and flexible representations of the intricate interactions taking place during battery degradation [6]. Furthermore, the attention mechanism enhances the model’s ability to focus on the most relevant features and time steps, improving the interpretability and predictive accuracy of the capacity prediction process [7-8].
In the proposed approach, the spatial attention is applied to voltage, current, and temperature measurements collected from each charging cycle. This module dynamically learns the relationships and dependencies among these features. In parallel, the temporal attention is applied across different timesteps within the data to enable the model to capture long-term dependencies and variations over the battery's operational history. The outputs from the spatial and temporal attention modules are then passed through a gated fusion mechanism. This ensures that the model can effectively balance the contributions from spatial feature correlations and temporal dynamics. Finally, the fused feature representations are passed through a series of fully connected layers to perform the final capacity prediction.
Our proposed method is validated using public benchmark datasets to demonstrate its robustness and effectiveness. Specifically, we employ the benchmark NASA battery dataset (Cells B5, B6, B7, and B18). To create different test scenarios, we use the last 80 cycles of each battery cell alternately as the test data while the cycling data of the other cells are used for training. Simulation results show that our model can consistently outperform traditional deep learning-based capacity prediction methods. The averaged RMSE (root mean squared error) values for capacity prediction of these cells, with similar model complexity of the proposed GNN, long short-term memory, and convolutional neural networks, are 0.0294, 0.1111, and 0.1810, respectively. These results confirm that our graph-based spatiotemporal attention framework can enhance the accuracy and reliability for the long-term lithium-ion battery capacity prediction.
References
[1] Koech, A. K., Mwandila, G., Mulolani, F., & Mwaanga, P. (2024). Lithium-ion battery fundamentals and exploration of cathode materials: A review. South African Journal of Chemical Engineering.
[2] Modekwe, G., Al-Wahaibi, S., & Lu, Q. (2024). Transformer-based Capacity Prediction for Lithium-ion Batteries with Data Augmentation. IFAC-PapersOnLine.
[3] Zhu, J. et al., (2022). Data-driven capacity estimation of commercial lithium-ion batteries from voltage relaxation. Nature Communications.
[4] Shi, Q., Zhao, L., Zhang, E., Xia, J., Li, H., Wang, K., & Jiang, K. (2023). The future capacity prediction using a hybrid data-driven approach and aging analysis of liquid metal batteries. Journal of Energy Storage.
[5] Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., & Yu, P. S. (2020). A comprehensive survey on graph neural networks. IEEE Transactions on Neural Networks and Learning Systems.
[6] Zhou, K. Q., Qin, Y., & Yuen, C. (2024). Graph neural network-based lithium-ion battery state of health estimation using partial discharging curve. Journal of Energy Storage.
[7] Zheng, C., Fan, X., Wang, C., & Qi, J. (2020, April). Gman: A graph multi-attention network for traffic prediction. Proceedings of the AAAI Conference on Artificial Intelligence.
[8] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems.