Data-Efficient Hybrid Sampling Method for Battery Health Estimation and Prediction - A comparative study of sampling methods for lithium-ion batteries using machine learning

Abstract

Abstract The rapid market adoption of electric vehicles (EVs) has been greatly driven by advances in lithium-ion batteries over the past years. As batteries are used under various load profiles in EVs, their capacity will decrease over time due to complex interactions of multiple degradation mechanisms. This necessitates advanced battery state-of-health (SOH) estimation and prediction. Existing methods rely on periodically sampled data for battery SOH estimation and prediction. However, with the ever-increasing measurements of battery data and complexity of algorithms, the limited hardware resources in the onboard battery management system (BMS) are strained. This makes the deployment of existing methods based on periodically sampled data challenging. More so, the research on the topic is limited and few sampling methods have been investigated, only validated on battery ageing datasets under static protocol cycling. The purpose of this thesis work is to develop data-efficient sampling methods for battery SOH estimation and prediction. The performance of two sampling methods has been evaluated on an open-source dataset for realistic EV driving profiles, and benchmarked to the periodic sampling method. Battery SOH estimations and predictions were produced using a Gaussian Process regression (GPR) as it provides a principled approach to handling uncertainties. To optimize the model, 29 features were manually extracted and 4 different feature sets were created based on the feature’s correlation with capacity. The experimental results showed that by using a 3-feature set with a hybrid of an event-based and periodic sampling method, a more accurate SOH prediction could be achieved while using significantly less data. Specifically, the proposed sampling method and feature set reduced RMSE by 45.28% and the required data used for inferring the model by 99.4%. To achieve even better results, improved features for the machine learning model should be investigated. The results of the work show that the method seems promising to alleviate the limited hardware resources of the BMS.