Patil, Dhruv (2026) Comparison of Forecasting Models for Short-Term Heat-Load Forecasting in District Heating Networks. Masters thesis, Universität Rostock.
Full text not available from this repository.Abstract
Reliable short-term heat-load forecasts are essential for model predictive control in district heating networks, since control actions depend on accurate estimates of future demand. However, heat-load forecasting is challenging because demand is influenced by ambient temperature, calendar effects, seasonal patterns, recent consumption behaviour, and changes in the active network structure over time. This thesis compares forecasting models for 24-hour-ahead heat-load prediction in a district heating network at 15-minute resolution. The study is based on data from the district heating network in Weil am Rhein covering the period from April 2020 to February 2026. Three external forecasting approaches were evaluated under identical conditions: Prophet, XGBoost, and a Prophet+XGBoost hybrid model. The results show that XGBoost achieved the best overall performance among the external models. The hybrid model improved over standalone Prophet, but did not outperform standalone XGBoost. In the benchmark comparison with Fraunhofer’s internal AEDL model and a simple Day-Before baseline, the normalised XGBoost configuration with temperature forecast feature achieved the strongest full-year performance across the main absolute and aggregate evaluation metrics, while AEDL achieved slightly better percentage-based error values. A more detailed seasonal analysis showed that active-building normalisation was beneficial during winter, whereas the non-normalised XGBoost configuration performed better during summer. The analysis of the first forecast step showed the same seasonal dependency, which is particularly relevant for control-oriented applications. The SHAP-based feature-importance analysis showed that recent consumption behaviour, especially rolling consumption statistics, was the dominant predictor, while future temperature information, active-building information, and calendar-related features also contributed to the forecasts. In addition, the weekly retraining setup improved forecasting performance compared with the fixed full-year evaluation. For the normalised XGBoost configuration, MAPE decreased from 11.71% to 8.93%, while AEDL improved from 10.53% to 9.31%. However, the seasonal behaviour remained visible. The normalised XGBoost configuration was still more suitable during winter, while the non-normalised configuration performed better during summer. Overall, the thesis shows that feature-engineered XGBoost models are suitable for short-term heat-load forecasting in the investigated district heating network and provide a promising basis for operational forecasting in control-oriented district heating applications.
| Item Type: | Thesis (Masters) |
|---|---|
| Subjects: | Autorenart > Studentische Arbeiten > Masterarbeit Autorenart > Studentische Arbeiten |
| Depositing User: | Dbis Admin |
| Date Deposited: | 06 Oct 2026 11:27 |
| Last Modified: | 06 Oct 2026 11:27 |
| URI: | https://eprints.dbis.informatik.uni-rostock.de/id/eprint/1161 |
Actions (login required)
 |
View Item |
