Evaluation of Forecasting Models for District Heating
Comparing a storage-aware surrogate with operational replay on identical forecast outputs to determine whether the two methods select the same model.
- Forecasting model families
- 6
- Random seeds
- 30
- Shared evaluation weeks
- 104
Data and methods
- The models were trained and tuned on data from 2014–2022, then tested on data from 2023–2024.
- The comparison covers six forecasting model families, 30 random seeds, and 104 common evaluation weeks.
- Each set of forecasts was assessed using both a storage-aware surrogate and an operational replay that accounts for unit commitment and state transitions.
Results
- HGB-r3 ranked sixth under surrogate evaluation but first under operational replay.
- When model selection was based on the first 52 weeks, the two methods selected different models for every one of the 30 random seeds.
- Over the subsequent 52 weeks, the selected models differed by an average of 5.414 weighted cost units per time step.
Status
The manuscript, “Forecast Model Selection for District Heating: Storage-Aware Surrogate Evaluation versus Operational Replay,” is currently under review.