Feature selection for specific prediction targets at the user level in a district heating network.

Area of Science:

• Energy Systems Engineering
• Data Science
• Sustainable Energy

Background:

• District heating networks are crucial for clean energy transitions but face challenges in efficiency and sustainability.
• Accurate modeling of district heating networks is essential for reducing energy losses and enhancing user comfort.
• Machine learning offers a powerful methodological approach to understand complex influencing factors and demand-side properties.

Purpose of the Study:

• To accelerate the application of machine learning in district heating network modeling.
• To generate knowledge on feature engineering and selection for predicting volume flow, supply, and return temperatures at the building level.
• To develop a systematic workflow for data acquisition and predictor selection.

Main Methods:

• Applied statistical and machine learning methods for feature engineering.
• Acquired data from a model region in northern Germany, including meteorological, behavioral, and operational parameters.
• Systematically selected the most relevant predictors for district heating network modeling.

Main Results:

• Temporal predictors and operational features from the infeed facility showed the highest relevance (15-20%).
• Outside air temperature, often key in heat load studies, was found to be of secondary relevance (6-10%) for the proposed prediction targets.
• Established specific interdependencies between various influencing factors and network performance.

Conclusions:

• The study provides a robust feature engineering and selection strategy for district heating network modeling.
• The findings offer crucial knowledge for efficient, machine learning-based modeling of these networks.
• This approach is a prerequisite for optimizing district heating systems towards greater sustainability and efficiency.