ShotAdpt: Few-Shot Early Prediction of Battery Cycle Life With Adaptive Feature Reduction

Accurate cycle life prediction of lithium-ion batteries using early-cycle data significantly enhances the safety, manufacturing, utilization, and development of batteries. However, existing methods are data-intensive and, therefore, face the challenge of sample scarcity under unseen operating conditions. Additionally, these methods require an operating-condition-specific feature reduction step to train the predictor, which leads to a generalization gap across unseen conditions. To address these issues, ShotAdpt is proposed as the first few-shot learning (FSL) framework that employs task-oriented adaptive feature reduction for early prediction of battery cycle life under unseen conditions with limited samples. It employs a meta-learning method to jointly optimize both feature reduction and prediction tasks, which is compatible with any gradient-based learning model. Extensive experimental results on a real battery dataset demonstrate ShotAdpt’s effectiveness in early prediction under unseen conditions, using only data from the first 100 cycles. Compared to state-of-the-art methods, ShotAdpt reduces the mean testing means absolute percentage error (MAPE) and root mean square error (RMSE) by factors of 13.04 and 12.69, respectively. Meanwhile, it maintains robust generalization performance across different unseen conditions and selected feature numbers. Furthermore, ShotAdpt achieves comparable prediction performance using only the first five cycles of data, whereas state-of-the-art methods require 100 cycles, demonstrating its potential for practical applications.