Prediction of Traction Battery Degradation in Electric Vehicles Using Ensemble Machine Learning Methods

Abstract

This paper addresses the problem of predicting degradation of traction lithium-ion batteries in electric vehicles based on real operational data. The objective is to develop and compare practical approaches to training machine learning models for forecasting battery remaining capacity (State of Health, SoH) using the open NASA Battery Dataset. The study implements a complete data processing pipeline: from loading and exploratory analysis to normalization, feature engineering, and training three candidate models: linear regression, random forest, and gradient boosting (XGBoost). Performance metrics include root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination $\mathrm{R}^{\mathrm{2}}$. Experiments demonstrate that ensemble methods (XGBoost) achieve the highest prediction accuracy $\left(\text{R M S E}=\text{1. 8 3} \%, \mathrm{R}^{\mathrm{2}}=\text{0. 9 6}\right)$ with acceptable inference time, making them suitable for integration into onboard battery monitoring systems. The practical significance of this work lies in the early detection of batteries with accelerated degradation to optimize warranty servicing, replacement planning, and reduction of total cost of ownership for electric vehicles.