Separating multiscale Battery dynamics and predicting multi-step ahead voltage simultaneously through a data-driven approach

TL;DR

This paper introduces a data-driven, health-aware battery model that separates fast and slow dynamics to improve multi-step ahead voltage prediction under aging conditions, verified on real EV driving data.

Contribution

It presents a novel sequence-to-sequence encoder-decoder model that infers degradation states and predicts battery voltage simultaneously, addressing complex aging mechanisms.

Findings

Effective separation of fast and slow battery dynamics.

Accurate multi-step voltage prediction under aging conditions.

Validated on real electric vehicle driving data.

Abstract

Accurate prediction of battery performance under various ageing conditions is necessary for reliable and stable battery operations. Due to complex battery degradation mechanisms, estimating the accurate ageing level and ageing-dependent battery dynamics is difficult. This work presents a health-aware battery model that is capable of separating fast dynamics from slowly varying states of degradation and state of charge (SOC). The method is based on a sequence-to-sequence learning-based encoder-decoder model, where the encoder infers the slowly varying states as the latent space variables in an unsupervised way, and the decoder provides health-aware multi-step ahead prediction conditioned on slowly varying states from the encoder. The proposed approach is verified on a Lithium-ion battery ageing dataset based on real driving profiles of electric vehicles.