(458g) Simultaneous Capacity and Power Fade Prediction of Lithium-Ion Batteries Using Multi-Task Learning

Lithium-ion batteries (LIBs) are widely used for electric vehicles (EVs), whose lifetime is affected by various factors (e.g. Li plating, solid electrolyte interphase formation, electrolyte depletion). State of energy (SoE) prediction of EVs is crucial for ensuring driver’s safety considering driving profiles, which is closely related to both capacity and power fades [1]. However, in the majority of the literature, capacity fade is only considered for predicting LIB’s cycling profiles and optimizing EV’s operating policies, without giving an attention to power fade.

Power fade is expected to play an important role, especially in fast charging and discharging cases, where capacity and power exhibit different nonlinear degradation patterns (e.g. different knee locations for capacity and power) [2]. For such cases, simultaneous prediction of capacity and power fade knees becomes crucial for accurate SoE prediction, and unveiling dominant degradation mechanisms for each fade. Feature engineering for LIB’s capacity fade prediction is a challenging problem, which still remains open, and it can be further complicated by considering both capacity and power fades. To the best of our knowledge, no research has predicted capacity and power knees simultaneously with solid input feature selection strategies.

Considering this gap, in this study, we make an attempt to identify important features which have strong capability in predicting both capacity and power fades. A public dataset [3] (i.e. Severson dataset) is employed to this end, which contains 124 LIB cells with various capacity and power fade behaviors. Selected features are extracted from direct health indexes (i.e. voltage, current, and temperature data). We initially consider a large set of feature candidates, which can be classified into two categories: temporal and inter-cycle features. Temporal features (e.g. constant voltage (CV) charging time and constant current charging time) are selected for input candidates, since many papers have illustrated their importance in capacity fade prediction (e.g. an increase in CV charging time can reflect internal resistance and charge capabilities changes) [4,5]. Additionally, inter-cycle features (e.g. average voltage changes from each cycle, power and energy values during charging and discharging) are utilized, since cyclic behaviors are shown to be crucial for LIB’s lifespan prediction [6].

From the feature candidates, hybrid feature selection methods are utilized to reduce model complexity by analyzing the first 100 cycle data [7], which incorporates the electrochemical knowledge, and selects the final features from filtering methods (e.g. using Pearson correlation coefficient). After the feature selection, multi-task learning is applied to predict both knees simultaneously. Linear regression models (e.g. elastic net and ordinary least squares) are used to enhance interpretability and regression performance.Our results show outstanding capacity and power knees prediction performance compared to the conventional methods [3,8]. Furthermore, detailed analysis for the selected features is provided, where different crucial features for predicting each knee are identified, linking with different major degradation mechanisms for each fade.

[1] Quade, K. L., Jöst, D., Sauer, D. U., & Li, W. (2023). Understanding the Energy Potential of Lithium‐Ion Batteries: Definition and Estimation of the State of Energy. Batteries & Supercaps, 6(8).

[2] Attia, P. M., Bills, A., Brosa Planella, F., Dechent, P., dos Reis, G., Dubarry, M., Gasper, P., Gilchrist, R., Greenbank, S., Howey, D., Liu, O., Khoo, E., Preger, Y., Soni, A., Sripad, S., Stefanopoulou, A. G., & Sulzer, V. (2022). Review—“Knees” in Lithium-Ion Battery Aging Trajectories. Journal of the Electrochemical Society, 169(6)

[3] Severson, K. A., Attia, P. M., Jin, N., Perkins, N., Jiang, B., Yang, Z., Chen, M. H., Aykol, M., Herring, P. K., Fraggedakis, D., Bazant, M. Z., Harris, S. J., Chueh, W. C., & Braatz, R. D. (2019). Data-driven prediction of battery cycle life before capacity degradation. Nature Energy, 4(5), 383–391.

[4] Ruan, H., He, H., Wei, Z., Quan, Z., & Li, Y. (2023). State of Health Estimation of Lithium-Ion Battery Based on Constant-Voltage Charging Reconstruction. IEEE Journal of Emerging and Selected Topics in Power Electronics, 11(4), 4393–4402.

[5] Chen, J., Chen, D., Han, X., Li, Z., Zhang, W., & Lai, C. S. (2023). State-of-Health Estimation of Lithium-Ion Battery Based on Constant Voltage Charging Duration. Batteries, 9(12), 565.

[6] Lee, J., & Lee, J. H. (2024). Simultaneous extraction of intra- and inter-cycle features for predicting lithium-ion battery’s knees using convolutional and recurrent neural networks. Applied Energy, 356, 122399.

[7] Zhao, B., Zhang, W., Zhang, Y., Zhang, C., Zhang, C., & Zhang, J. (2024). Research on the remaining useful life prediction method for lithium-ion batteries by fusion of feature engineering and deep learning. Applied Energy, 358, 122325.

[8] Saxena, S., Ward, L., Kubal, J., Lu, W., Babinec, S., & Paulson, N. (2022). A convolutional neural network model for battery capacity fade curve prediction using early life data. Journal of Power Sources, 542, 231736.