Early prediction of battery life by learning from both time-series and histogram data

Paper i proceeding, 2023

Due to dynamic operating conditions, random user behaviors, and cell-to-cell variations, accurately predicting battery life is challenging, especially using information from only a few early cycles. This work proposes a data-driven battery early prediction pipeline using both time-series, measurement-related features, and usage-related histogram features. We
first investigate the prediction performance of using these two feature sources individually, then two methods of systematically combining these two feature sources are devised. Additionally, four machine learning algorithms with different characteristics are applied to compare their performances on battery prognostic problems. We show that the prediction accuracy of using these two feature sources individually is comparable. Moreover, a systematic combination of these two features considerably improves the prediction performance in terms of accuracy and robustness, achieving excellent prediction results with a root mean square error of around 150 cycles using only the first 100 cycle’s data. Finally, experimental data of different cell types and cycling conditions are used to verify the developed method’s effectiveness and generality.