Weighted Conformal Prediction is a significant advancement in predicting thermal volatility in electric vehicle (EV) powertrains, particularly in the high-performance motorsport sector. On July 2, 2026, Varshith Roy Kotla published a study that addresses the challenges of lab-to-track thermal transfer, revealing that conventional models often fail under real-world conditions.
The study highlights the difficulty in observing internal temperatures of EV powertrains outside laboratory settings. Models calibrated on lab drive cycles do not perform well when faced with actual driving loads, leading to unreliable predictions. To tackle this issue, the research employs conformal prediction, which provides distribution-free uncertainty bounds, thereby improving prediction accuracy.
Implementation of Ensemble Batch Prediction Intervals
Kotla's research implements Ensemble Batch Prediction Intervals (EnbPI), a method designed for autocorrelated time series, as introduced by Xu and Xie in 2021. The model is calibrated using real data from a CALCE lithium-ion cycler, specifically A123 SP20 cells under a FUDS profile. The study evaluates the model under a measured covariate shift, using a second CALCE test condition during the US06 Highway Driving Schedule at 45°C.
Results show that the unweighted EnbPI bound, which achieved a nominal 95% coverage in distribution, dropped to only 70.13% empirical coverage under the real-world conditions. This significant decline underscores the need for improved methodologies in thermal prediction.
Advancements with Weighted EnbPI Procedure
The study introduces a weighted EnbPI procedure that combines the ensemble residuals from EnbPI with density-ratio weighting, as suggested by Tibshirani et al. in 2019. This innovative approach estimates the density ratio through a probabilistic domain classifier, resulting in an improved coverage of 72.42%.
While this represents a modest improvement, Kotla emphasizes that it does not completely resolve the challenges associated with lab-to-track thermal transfer. The calibrated model was further applied to real 2023 Formula 1 telemetry data from races at Monza and Silverstone, focusing on driver VER. However, due to the lack of an internal thermal channel in public telemetry, the study reports only unsupervised flag rates of 65.6% at Monza and 58.0% at Silverstone, both significantly exceeding the 5% in-distribution base rate.
Conclusions on Conformal Domain Adaptation
Kotla concludes that conformal domain adaptation serves as a promising yet partially resolved tool for addressing the lab-to-track thermal transfer problem. The research details specific areas where the methodology falls short, paving the way for future studies to build upon these findings.
- Publication Date: July 2, 2026
- Author: Varshith Roy Kotla
- Coverage Improvement: From 70.13% to 72.42%
- Formula 1 Telemetry Flag Rates: 65.6% (Monza), 58.0% (Silverstone)
🤖 This article was rewritten by Feed and Figures' editorial AI from a report originally published by arXiv Machine Learning. Facts and quotes are preserved from the original; the rewrite focuses on clarity and structure. For the unedited original, see the source link below.
