Evaluation of Forecasting Models for District Heating
State Grid Zhejiang Electric Power Research Institute · Research experience · Jul 2025 – Apr 2026
Hourly heat-load forecasting 24 hours ahead, with model selection based on forecast accuracy and operating cost in rolling dispatch.
- Candidate models
- 7
- Forecast horizon
- 24 h
- Status
- SCI manuscript under review
Research
Prepared historical district-heating data to study how forecast-model selection affects coordinated operation of CHP, thermal storage and auxiliary heat sources.
Data and methods
- Aligned timestamps, handled missing values and built features for hourly heat-load forecasts.
- Compared seven candidate models based on HGB, MLP and TCN, examining training objectives and bias-correction methods.
- Connected forecasts to rolling dispatch and evaluated models using forecast errors, equipment constraints and operating costs.
Results
- Completed approximately 96,000 hours of forecasting experiments and built a training and operational-evaluation framework.
- In baseline experiments, model selection using full operational evaluation reduced operating cost by 5.34% versus error-based selection and 3.78% versus simplified operational evaluation.
Status
One related SCI journal manuscript is under review.