Battery Charge Curve Prediction via Feature Extraction and Supervised Machine Learning

Area of Science:

• Materials Science
• Electrochemistry
• Battery Technology

Background:

• Real-time battery state monitoring is crucial for safe and reliable operation of battery-powered devices.
• Predicting battery performance over its lifespan requires accurate state estimation.
• Current methods may require extensive data or long monitoring periods.

Purpose of the Study:

• To develop a methodology for predicting the entire battery charge curve using limited initial data.
• To enable fast, onboard health status monitoring and estimation for batteries.
• To assess the generalization capability of the prediction method across different battery chemistries.

Main Methods:

• Collected 10,066 charge curves from LiNiO2-based batteries at a constant C-rate.
• Employed a combination of feature extraction and multiple linear regression.
• Validated the methodology on open-access datasets of LiCoO2-based batteries.

Main Results:

• Accurately predicted entire battery charge curves with less than 2% error using only 10% of the data for LiNiO2-based batteries.
• Achieved approximately 2% prediction error for LiCoO2-based batteries using only 5% of the charge curve data.
• Demonstrated the generalization of the prediction methodology across different battery chemistries.

Conclusions:

• The developed method enables accurate prediction of battery cycling curves with minimal input data.
• This approach facilitates rapid onboard health monitoring for practical battery applications.
• The methodology shows strong generalization, applicable to various battery types.