Machine Learning-based Online State-of-Health Estimation of Electric Vehicle Batteries: Artificial Intelligence Applied to Battery Management Systems
Translated title
Machine Learning-based Online State-of-Health Estimation of Electric Vehicle Batteries
Author
Barragan Moreno, Alberto
Term
4. term
Education
Publication year
2021
Abstract
Efterhånden som elbiler bliver mere udbredte, er det vigtigt at kende batteriets helbredstilstand (State of Health, SoH) – et mål for hvor meget kapacitet der er tilbage, og hvor meget den indre modstand er steget. Dette speciale anvender maskinlæring til at estimere SoH i realtid under kørsel, på tværs af forskellige litium‑ion‑kemier og praktiske brugssituationer. Efter en gennemgang af, hvordan celler ældes, og hvilke faktorer der påvirker processen, udvikles tre datadrevne metoder: (1) at udlede SoH ud fra spændingsmålinger ved faste ladetilstande (State of Charge, SoC); (2) at bruge ladningsgradienter – hvordan den lagrede ladning ændrer sig mellem bestemte spændingspunkter – som helbredsindikatorer; og (3) at forudsige kapacitets‑ og impedansbaseret SoH ud fra et lille sæt impedansmålinger (et mål relateret til indre elektrisk modstand). Metoderne afprøves i realistiske scenarier, herunder varierende temperaturer, forskellige aldringsforløb og opbevaringsaldring. De opnår høj nøjagtighed med middel absolut fejl (MAE) helt ned til 0,4 % for fremtidige forudsigelser, 0,5 % ved flere temperaturer, 0,4 % ved realistisk, varieret aldring og 0,2 % ved opbevaringsaldring. Resultaterne viser, at metoderne er praktiske og pålidelige til online SoH‑estimering i virkelige elbiler.
As electric vehicles become more common, it is important to know a battery’s state of health (SoH)—a measure of how much capacity remains and how much internal resistance has increased. This thesis applies machine learning to estimate SoH in real time during vehicle operation, across different lithium‑ion chemistries and practical use conditions. After examining how cells age and what drives that aging, three data‑driven methods are developed: (1) inferring SoH from voltage measurements taken at fixed state‑of‑charge (SoC) levels; (2) using charge gradients—how the stored charge changes between specific voltage points—as indicators of health; and (3) predicting capacity‑ and impedance‑based SoH from a small set of impedance measurements (a measure related to internal electrical resistance). The methods are evaluated under realistic scenarios, including varying temperatures, diverse aging profiles, and storage aging. They achieve high accuracy, with mean absolute error (MAE) as low as 0.4% for future SoH predictions, 0.5% across multiple temperatures, 0.4% under diverse, realistic aging, and 0.2% for storage aging. These results show that the approaches are practical and reliable for online SoH estimation in real‑world EVs.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
Documents