A Lightweight Multitask Learning Framework for Battery Degradation Prediction With Limited Data

Accurate prediction of nonlinear aging is paramount for ensuring the reliable operation and timely maintenance of battery energy storage systems. However, stable battery aging prediction remains challenging due to diverse battery types, aging mechanisms and working conditions. This study explores the feasibility of utilizing limited data to comprehensively evaluate battery degradation through advanced deep learning techniques. A lightweight multi-task learning (LMTL) model is proposed to simultaneously predict the capacity knee-point (KP) and remaining useful life (RUL). It employs the partial voltage sequences from limited number of cycles, enabling flexible adaptation to various stages of battery life in an end-to-end framework. To enhance deployability, projection compression technology is integrated, reducing the model size by over 90% without compromising accuracy. A total of 227 batteries, covering three types of cathode chemistries, are utilized to validate the proposed method. Experimental results indicate that the mean absolute percentage errors (MAPE) of RUL and KP are within 4.1% and 5.9%, respectively, outperforming traditional approaches. Transfer learning further improves generalization, reducing average prediction errors by over 25% with only 10% of the target domain data. In addition, the effects of input window size, partial sequence length, and applicability to battery packs are thoroughly investigated, offering practical guidance for real-world deployment. By minimizing the testing and computational costs, the proposed method offers a robust and flexible solution for applications such as battery design, residual value assessment, and performance maintenance.