Real World Short Term District Heating Load Forecasting A Machine Learning Benchmark Against an Operational Utility Model

In urban district-heating networks, operators plan unit dispatch, storage operation, and fuel procurement on the basis of short-term heat-demand forecasts. At the Stadtwerke Karlsruhe GmbH (Germany), this forecast has been generated since 2017 by a compact gray-box model that adjusts the load profile of a reference-day using a 72 h temperature forecast and a weekday-dependent calendar correction. We benchmark this operational reference model against machine-learning models (namely ordinary least-squares, Ridge, and ElasticNetregression, a multi-layer perceptron, XGBoost, LightGBM, and the tabular foundation models TabPFN and TabFM). On the 2025/26 test season, the best-performing model, TabPFN, reduces the mean absolute error from 14.46 MW of the operational reference model to 10.69 MW, a reduction of 26 %. TabPFN obtains these results without a task-specific weight update, using the training seasons only as in-context examples. In comparison, the regularized linear model ElasticNet reaches 10.78 MW and thus achieves 97.6 % of the improve-ment achieved by TabPFN, while remaining interpretable: its coefficients reveal a plausible learned calendar correction and temperature forecast dependency. For similar applications, the reported results can serve as a guide to trade implementation effort against accuracy and explainability.