New Conformal Prediction Method Improves EV Powertrain Thermal Modeling

Predicting the thermal behavior of electric vehicle powertrains is complex, especially when trying to apply models trained in controlled lab environments to the unpredictable demands of real-world racing. Internal temperatures are often unobservable outside the lab, leading to significant model degradation when deployed on track. This research explores using conformal prediction, specifically an adapted Ensemble Batch Prediction Intervals (EnbPI) method, to provide robust uncertainty bounds for these thermal predictions. The standard EnbPI method, while effective in-distribution, showed a substantial drop in predictive coverage when faced with real-world covariate shifts. To counter this, the team introduced a weighted EnbPI procedure that combines ensemble residuals with density-ratio weighting. This enhanced approach demonstrated a modest but consistent improvement in recovering predictive coverage under challenging real-world conditions, though it doesn't fully resolve the problem. The calibrated model was also applied to Formula 1 telemetry as an unsupervised diagnostic, highlighting its potential for identifying anomalous thermal events.