Toward explainable heat load patterns prediction for district heating.

Area of Science:

• Energy Systems Engineering
• Artificial Intelligence in Energy
• Thermal Engineering

Background:

• District heating networks are crucial for urban thermal energy supply.
• Optimizing heat networks requires understanding user heat consumption patterns.
• Existing research often overlooks detailed heat usage analysis or operates on a limited scale.

Purpose of the Study:

• To develop and evaluate a data-driven approach for analyzing and predicting heat load in a district heating network.
• To address the gap in large-scale heat usage profile analysis.
• To compare the efficacy of different machine learning algorithms for heat load forecasting.

Main Methods:

• Utilized eight heating seasons of data from a cogeneration district heating (DH) plant in Cheongju, Korea.
• Developed predictive models using supervised machine learning (ML): Support Vector Regression (SVR), Boosting algorithms, and Multilayer Perceptron (MLP).
• Input variables included weather data, holiday information, and historical hourly heat load; model performance was assessed using varying training data sizes.

Main Results:

• Boosting algorithms, specifically XGBoost, demonstrated lower prediction errors compared to SVR and MLP.
• The study identified XGBoost as a highly suitable ML algorithm for accurate heat load forecasting.
• Explainable AI techniques were employed to interpret model behavior and variable importance.

Conclusions:

• Machine learning, particularly boosting algorithms like XGBoost, offers a robust solution for accurate heat load prediction in district heating networks.
• The findings provide valuable insights for optimizing the operational efficiency and capacity planning of district heating systems.
• This data-driven approach, supported by explainable AI, enhances the understanding and management of urban thermal energy distribution.