A Machine Learning Framework for Battery Thermal Rise Prediction During EV Fast Charging

Abstract

Electric vehicle fast charging presents a three-way conflict between charging speed, battery longevity, and thermal safety that fixed constant-current/constant-voltage protocols cannot resolve adaptively. This thesis develops a machine learning framework to predict lithium-ion battery temperature rise above ambient during fast charging, providing the thermal prediction layer needed to support future adaptive charging controllers. The study uses 2,813 metadata-confirmed charge sessions from 34 batteries, totalling 6,214,672 cleaned charge rows. Each session contributes 6 representative rows, giving 16,738 modelling rows. The prediction target, Temperature_Rise_C, is defined as battery surface temperature minus ambient temperature. Thirteen machine-learning algorithms spanning linear, tree-based ensemble, instance-based, kernel, and neural network paradigms are evaluated against a dummy mean baseline and a linear regression baseline under three validation schemes: row-random, session-holdout, and battery-holdout. Against the dummy mean baseline (RMSE = 2.862 °C) and linear regression baseline (RMSE = 2.312 °C), Random Forest achieves the best session-holdout result: R² = 0.8466, RMSE = 1.121 °C, representing improvements of 60.8% and 51.5% respectively. Under battery-holdout validation, K-Nearest Neighbors leads at R² = 0.4871 and RMSE = 2.534 °C. Voltage state and state-of-charge proxy are the two strongest predictors at 18.3% and 16.5% feature importance. Cross-battery generalisation is identified as the primary barrier to deployment. The study supplies a validated lab-scale prediction layer that establishes the thermal estimation foundation required for future adaptive charging controllers, current-schedule optimisation, and battery-health-aware fast charging systems.