Enhanced Gaussian process dynamical modeling for battery health status forecasting

Abstract

Monitoring the state-of-health of Li-ion batteries is a critical component of battery management systems in electric vehicles. A large number of feature-based machine-learning methods have been introduced in the last decade to improve the accuracy of predictions of the state-of-health and end-of-life, especially early in the lifetime of the battery stack. Unless multiple battery data sets are used for direct and crude predictions of the end-of-life, however, such an approach is infeasible since the features are not known for future cycles. In this study a new nonlinear state-space model that can overcome this limitation is introduced. The powerful Gaussian process dynamical model is extended by generalizing the covariance structure, and therefore permitting more flexible models for the observables and latent variables. The model is further enhanced with transfer learning, to yield accurate early predictions of the future state-of-health of Li-ion batteries up to end-of-life. Experiments conducted on two of the NASA Ames Battery data sets and the Oxford Battery Degradation data set demonstrate the accuracy and superiority of the new model over state-of-the-art benchmarks algorithms, including supervised Gaussian process models, deep convolutional networks, recurrent networks and support vector regression. The root mean square error is reduced by up to 43% on the NASA data sets and by up to 54% on the Oxford data set.