Two-stage surrogate modeling framework for heat demand and indoor temperature prediction in district heating systems

Accurate forecasting of heating demand and indoor thermal comfort is essential for optimizing the operation of District Heating Systems (DHS), particularly when detailed measurements on the consumer side are unavailable. This paper proposes two-stage surrogate modeling approach, integrating state-of-the-art Machine Learning methods to separately model heat supply and indoor temperature dynamics. In the first stage, a Transformer-based model is employed to forecast 24-h profiles of supplied heating energy using SCADA and weather data. In second stage, a gradient-boosted decision tree model predicts indoor temperatures based on forecasted heat supply, ambient temperature, and temporal features. The surrogate model emulates the behavior of the building subsystem without requiring direct access to indoor measurements during deployment. The methodology was validated on multi-year operational data from a real-world DHS. Results show high sequence forecasting accuracy (MAE = 20.2 kWh), with sub-degree mean absolute error for room temperature predictions (MAE = 0.26 °C). Mean sequence accuracy of and accuracy of regression model was. In a demonstration of the approach, the paper also introduces a diagnostic interface for interactive exploration of prediction outputs, supporting operational insight. The proposed approach offers a scalable and practical tool for data-driven demand-side planning and contributes to the digital transformation of DHS operations.