EV motorsport thermal prediction gets 2.3% boost from weighted conformal method

Varshith Roy Kotla's research tackles a critical gap in EV motorsport: lab models for thermal prediction fail under real-world loads. The paper, submitted to arXiv (cs.LG), uses conformal prediction – specifically Ensemble Batch Prediction Intervals (EnbPI) – to quantify uncertainty without distributional assumptions. Calibrated on CALCE lithium-ion cycler data (FUDS profile, A123 SP20 cells), the unweighted EnbPI achieved its nominal 95% coverage in-distribution. However, under a genuine covariate shift (US06 Highway Driving Schedule at 45°C), coverage plummeted to 70.13%. To address this, the author introduces a weighted EnbPI that combines ensemble residuals with density-ratio weighting (Tibshirani et al., 2019), estimating ratios via a probabilistic domain classifier. This recovers coverage to 72.42% – a modest, honestly-reported improvement, not a complete fix.

The study also applies the calibrated model to real 2023 Formula 1 telemetry (driver VER at Monza and Silverstone) as an unsupervised out-of-distribution diagnostic. Since no internal thermal channel exists in public trackside data, only unsupervised flag rates are reported: 65.6% at Monza and 58.0% at Silverstone, far above the 5% in-distribution base rate. Notably, flags showed inconsistent associations with braking/DRS zones. Kotla concludes that conformal domain adaptation is promising but only partially solved, detailing exactly where it falls short. The paper includes 8 pages, 3 tables, and a GitHub link for code and data.