An attention‐based synthetic battery data augmentation technique to overcome limited dataset challenges

Abstract

Large‐scale global efforts for electrifying different modes of transportation have fueled the need for energy‐dense storage systems. Lithium‐ion batteries have shown promise to be a favourable technology to power pure electric vehicles, off‐board storage systems for microgrids and other hybrid vehicle applications. Accurate estimation of vital battery parameters, such as state‐of‐charge, enable the user to predict remaining useful life of lithium‐ion batteries. Although multiple data‐driven techniques over the years have improved prediction accuracy, lack of high‐quality data continues to be a challenge which has been largely ignored in literature. Commercial battery manufacturers fail to release their datasets due to privacy concerns thereby worsening this crisis. The limited data challenge has been further aggravated due to advent of novel charging techniques, such as pulsed discharge and constant‐temperature charging profile. This article aims to resolve this data scarcity issue by employing an attention‐based technique to generate reliable synthetic data. The generated data will greatly assist researchers to develop reliable battery ageing prediction models. The approach followed in this article consumes very less computational resources as well.

Data Augmentation and Feature Selection for the Prediction of the State of Charge of Lithium-Ion Batteries Using Artificial Neural Networks

“…A possible solution to overcome these problems is the usage of artificially augmented data. In the last few years, it has been shown that data augmentation techniques can lead to improved results for ML models, thus making it possible to successfully tackle the problem of limited datasets [29][30][31].…”