Remaining useful-life prediction of lithium battery based on neural-network ensemble via conditional variational autoencoder

Abstract

Ensemble learning using deep neural networks has become prevalent in predicting the Remaining Useful Life (RUL) of Lithium Batteries (LiBs). However, owing to the predominant linearity of ensemble learning, capturing nonlinear relationships among base learners remains a persistent challenge. This study presents an RUL-prediction method for LiBs based on a neural-network ensemble via a Conditional Variational Autoencoder (CVAE). The proposed method serves as a nonlinear ensemble learning method and promises enhanced prediction performance while maintaining ease of implementation. The methodology entails several key steps. First, data smoothing is conducted via local weighted linear regression. Subsequently, a preliminary linear-ensemble phase is executed through an attention mechanism, which filters out extraneous information in the features and bolsters the importance of valid features. Subsequently, a nonlinear ensemble is accomplished by utilizing the CVAE, with truth labels serving as conditions. Finally, the efficacy of the proposed method is substantiated through experimentation, demonstrating its superior performance compared to the candidate methods.