A preprint comparing forecasting models on two German district-heating networks found a split result: Chronos-2 had the lower aggregate full-year error, while TabPFN-TS produced uncertainty intervals closer to their stated 80% coverage. TabPFN-TS also remained among the leading models when the comparison was transferred to the second network.
The study tested whether TabPFN-TS could make both point forecasts—single estimates of heat load—and probabilistic forecasts that show a range of possible outcomes, using recent observations and known future weather covariates. In the zero-shot setup, the model was used without task-specific gradient-based training or manual preprocessing; its 50th quantile supplied the point forecast, while its other quantiles were used for probabilistic evaluation.
A narrow gap on overall accuracy
On the main dataset, TabPFN-TS recorded a CVRMSE of 13.06%, compared with 12.48% for Chronos-2. CVRMSE was the study’s aggregate point-forecast error measure, so the lower figure went to Chronos-2. The study’s daily-rank comparison nevertheless kept the two models within its critical-difference threshold.
The model ordering was broadly preserved in transfer validation. Chronos-2, TabPFN-TS and TFT were the three strongest models in both networks, with Chronos-2 recording the lowest aggregate error; TabPFN-TS remained ahead of the trained AutoGluon baselines.
The forecast setup mattered
The selected TabPFN-TS setup used hourly forecasts 24 hours ahead, a 12-week rolling context and ambient temperature as the only weather covariate. Longer context windows did not improve accuracy.
At 15-minute resolution, the best four-hour setup used four weeks of context and had a CVRMSE of 14.04%, versus 15.71% for the best 24-hour setup. The reported gap was 1.67 percentage points, or 10.6% relative.
Better-calibrated ranges, but weather still matters
The advantage shifted when the researchers assessed uncertainty. TabPFN-TS had lower MACE, the study’s calibration measure, while Chronos-2 had the lower CRPS, the study’s overall probabilistic score, and tended to be slightly overconfident.
For nominal 80% intervals, TabPFN-TS covered 80.50% of observed outcomes in Munich, against 77.73% for Chronos-2. In Flensburg, the corresponding figures were 79.82% and 78.19%. TabPFN-TS was therefore closer to the target in both networks.
The full-year comparison used realized future temperature, while a sensitivity test substituted archived 24-hour-ahead temperature forecasts in Munich. That change corresponded to CVRMSE increases of 8.3% for TabPFN-TS and 9.5% for Chronos-2, with Chronos-2 slightly more sensitive in that test.
A second design traded detail for speed
The researchers also evaluated a Multi-Resolution Residual-Correction Forecaster, or MRRC, derived from the diagnostics. Compared with the High Frequency & High Resolution setup, MRRC produced similar short-term results, a reported 2.7% lower 12-hour integrated error, 4.1% lower bias magnitude and 48.7% lower RTF, the study’s runtime measure.
The authors present MRRC as a balance between short-term trajectory accuracy, longer-horizon energy planning and computational cost, rather than as evidence that it is universally superior.
A promising comparison, with clear boundaries
The evaluation was retrospective and covered only two German networks: a comparatively small, expanding Munich network and a substantially larger, comparatively stable Flensburg network. That limited scope means the results cannot by themselves establish how the models would perform beyond the two tested cases.
The study does not show that either model improves network operation, costs, emissions or storage scheduling, and it does not establish that MRRC is superior in every setting. No confidence intervals or inferential significance tests were reported for the comparisons.
Before operational use, the approach needs testing on additional networks and with realistic weather forecasts over longer horizons. The supplied analysis also leaves open how well the models can adapt to gradual changes in system behavior, known as concept drift.
The paper’s status
The work is an arXiv version 1 preprint dated 20 August 2026. Experiment code is available on GitHub and Zenodo; the Munich dataset is restricted by third-party conditions, while the Flensburg dataset is publicly available.
Paper data and sources
Original title: Systematic Evaluation of TabPFN-TS for Zero-Shot Probabilistic Heat Load Forecasting in District Heating Networks
Authors: Ben Spoek, Karim K. Ben Hicham, Kai Derzsi et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text