TabPFN-TS Evaluated for Zero-Shot Heat Load Forecasting

Reliable heat load forecasts are essential for efficient operation of district heating networks, but traditional methods require retraining models when network conditions change. Zero-shot time-series foundation models offer a promising alternative by adapting at inference time from recent observations. This research systematically evaluates TabPFN-TS, a model pretrained on synthetic data, for probabilistic heat load forecasting in district heating networks. The study compares TabPFN-TS against established time-series foundation models like Chronos-2 and other trained machine learning baselines. Researchers analyzed various factors including covariate choice, context length, temporal resolution, and prediction horizon. A configuration of hourly 24-hour forecasting with a 12-week rolling context and ambient temperature proved to be a parsimonious high-performing setup. Results indicate that TabPFN-TS achieves deterministic accuracy close to Chronos-2 (CVRMSE values of 13.06% vs. 12.48%) and shows superior empirical calibration. While Chronos-2 had the lowest aggregate full-year error, TabPFN-TS's performance, despite synthetic pretraining, highlights its potential for transferability and adaptability in dynamic district heating environments. The findings also motivate a Multi-Resolution Residual-Correction Forecaster for improved long-horizon planning.