Deep learning-based prognostic approach for lithium-ion batteries with adaptive time-series prediction and on-line validation

Abstract

Abstract Prognostics for lithium-ion batteries is very critical in many industrial applications, and accurate prediction of battery state of health (SOH) is of great importance for health management. This paper proposes a novel deep learning-based prognostic method for lithium-ion batteries with on-line validation. An effective variant of recurrent neural network, i.e. long short-term memory structure, is used with variable input dimension, that facilitates network training with additional labeled samples. Adaptive time-series predictions are carried out for prognostics. An on-line validation method is further proposed for parameter optimization in real time based on the available system information, which allows for continuous model improvement. Experiments on a popular lithium-ion battery dataset are implemented to validate the effectiveness and superiority of the proposed method. The experimental results show the prognostic performances are promising both for the multi-steps-ahead predictions and long-horizon SOH estimations.