Generating high‐fidelity synthetic battery parameter data: Solving sparse dataset challenges

Abstract

Global burgeoning pollution levels have led to massive efforts being made to electrify all modes of transportation in the coming decade. Most of the Electric Vehicles (EVs) in the automotive domain are powered by Lithium-ion batteries (LIBs). Steady increase in number of EVs has led to generation of enormous amounts of data. Most of this data is related to the usage of LIBs and its parameters such as voltage, current, and temperature values. Recent data-driven techniques possess the ability to extract meaningful insights from this data such as predicting the range of an Electric Vehicle or estimating State-of-Health of the batteries. Improvements in these models have been severely restricted due to limited access to proprietary experimental data and privacy concerns. This paper aims to introduce a Generative Adversarial Network based approach to produce high-fidelity synthetic data to overcome the limitation of limited data. The synthetic data produced can be used for training neural networks to improve the accuracy of battery parameter predictions. The objectives of this research work are threefold; first we try to produce high-fidelity synthetic time series data, secondly, heterogeneity in generated data is maintained, and thirdly, we test whether generated synthetic data enables improvements in accurate State-of-Charge estimation.