Interpretable Battery Lifetime Prediction Using Early Degradation Data

Licentiate thesis, 2023

Battery lifetime prediction using early degradation data is crucial for optimizing the lifecycle management of batteries from cradle to grave, one example is the management of an increasing number of batteries at the end of their first lives at lower economic and technical risk.

In this thesis, we first introduce quantile regression forests (QRF) model to provide both cycle life point prediction and range prediction with uncertainty quantified as the width of the prediction interval. Then two model-agnostic methods are employed to interpret the learned QRF model. Additionally, a machine learning pipeline is proposed to produce the best model among commonly-used machine learning models reported in the battery literature for battery cycle life early prediction. The experimental results illustrate that the QRF model provides the best range prediction performance using a relatively small lab dataset, thanks to its advantage of not assuming any specific distribution of cycle life. Moreover, the two most important input features are identified and their quantitative effect on predicted cycle life is investigated. Furthermore, a generalized capacity knee identification algorithm is developed to identify capacity knee and capacity knee-onset on the capacity fade curve. The proposed knee identification algorithm successfully identifies both the knee and knee-onset on synthetic degradation data as well as experimental degradation data of two chemistry types.

In summary, the learned QRF model can facilitate decision-making under uncertainty by providing more information about cycle life prediction than single point prediction alone, for example, selecting a high-cycle-life fast-charging protocol. The two model-agnostic interpretation methods can be easily applied to other data-driven methods with the aim of identifying important features and revealing the battery degradation process. Lastly, the proposed capacity knee identification algorithm can contribute to a successful second-life battery market from multiple aspects.