Do Foundation Models Generalize to Real-World EV Fleets? A 1.1M-Drive Benchmark

Patrick Ahrend, Stefan Pfund, Felix Divo

Presentation Attendance: No, we cannot present in-person

Keywords: mixture of experts, time series foundation models, electric vehicles, thermal management, time series clustering

TL;DR: A Cluster-Aware Mixture of Experts outperforms zero-shot foundation models and supervised baselines on 1.1M real-world EV driving sequences by routing inputs to specialized LSTM experts via learned SoftDTW-based gating.

Abstract:

Time series foundation models (TSFMs) promise general-purpose forecasting, yet
their effectiveness on large-scale industrial data with physical heterogeneity remains
underexplored. We benchmark three zero-shot TSFMs and three supervised models
against a Cluster-Aware Mixture of Experts on 1.1 million real-world electric
vehicle driving sequences. The Cluster-Aware approach routes inputs to specialized
LSTM experts based on Soft Dynamic Time Warping clusters, reducing mean
absolute error by 14.7% over a global LSTM baseline and outperforming all
other models. Evaluation on an unseen vehicle model confirms robust transfer
with only 8.1% increase in error. Our results suggest that while TSFMs deliver
competitive zero-shot performance, domain-informed supervised specialization
remains advantageous on this heterogeneous industrial dataset.

Track: Industry and Application Track (max 2 pages)

Submission Number: 105