A Hybrid Machine Learning Model for Battery Cycle Life Prediction with Early Cycle Data

Abstract

Accurate prediction of lithium-ion (Li-ion) battery cycle life using early cycle data is a challenging task as the capacity fade resulting from the nonlinear degradation process leads to a negligible loss of capacity in early cycles but is accelerated when approaching the end of life. To address this challenge, we propose a hybrid machine learning model that combines a shallow learning model, relevance vector machine (RVM), and a deep learning model, convolutional neural network (CNN). RVM is employed to generate artificial cells with high cycle lives, expanding the original training dataset (i.e., data augmentation). The expanded training dataset then serves as the input-output pairs used to train the CNN model. CNN first learns the locally-invariant features from the input data and then makes full use of these features to predict the cycle life of a Li-ion battery cell. We evaluate the performance of the proposed hybrid machine learning (RVM+CNN) model on two test datasets consisting of 83 cells with widely varying cycle lives ranging from 150 to 2300 cycles. The RVM+CNN model produces higher cycle life prediction accuracy on both datasets than three other machine learning and deep learning methods.