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

With the challenge of district heating network transition as part of the global objective of clean energy, machine learning provides a methodological approach for understanding the relationships between various influencing factors and demand-side properties of district heating networks, which is decisive for reducing losses, enhancing sustainability, and guaranteeing residential comfort. This work focuses on accelerating the application of modern machine learning methods to modeling district heating networks by generating knowledge on feature engineering and selection for newly suggested prediction targets, namely volume flow, supply, and return temperatures, directly at the building level. A systematic workflow for data acquisition, feature engineering, and selecting the most relevant predictors is presented. For this, statistical and machine learning methods are applied to engineer respective features and establish specific interdependencies, including meteorological conditions, human behavioral patterns, and operational parameters, based on a model region in northern Germany. The qualitative results indicate that the highest impact is for temporal predictors and operational features derived from the infeed facility's data, i.e., approximately 15 to 20% of the total predictor relevance. In comparison to studies targeting the heat load and suggesting outside air temperature as the most relevant predictor, it was found that for the herein proposed prediction targets, this feature is of secondary relevance (roughly 6-10%). The findings of this study provide a feature engineering and selection strategy, as well as relevant knowledge gain, which is a prerequisite for efficient modeling of district heating networks based on machine learning in the future.