Dynamic user-level parameter modeling in heat networks with deep learning Long Short-Term Memory

Providing access to user-level information in heat networks is crucial for optimizing control and facilitating the transition to carbon neutrality. Machine learning can automate model development in this area. While models for heat demand have been successfully applied to multi-apartment homes, buildings such as single-family homes pose challenges due to more complex data dynamics. This study aims to develop a Long Short-Term Memory modeling framework that captures high user-level data dynamics, focusing on the parameters of volume flow, supply, and return temperatures, beyond just heat demand. Sensitivity and comparative analysis of modeling errors evaluate data pre-preparation, hyperparameter tuning, and sampling strategies. Initially, data from 15 users are analyzed and rationalized. The model is then scaled up by increasing the number of substations to validate findings. Results show a mean absolute error of 0.0612 for heat demand prediction over 24 h in the upscaled model. Challenges such as the impacts of human behavior at the respective buildings and missing data are discussed. These insights can enhance automation of user-level parameter modeling, improving heat network control and supporting sustainable heat supply. • LSTM models automatically developed for highly dynamic user-level heat network parameters. • Enhanced model performance through sensitivity analysis during data pre-preparation. • Hyperparameter optimization accelerated by analysis of initially chosen intervals. • Improved data selection by analyzing seasonality and data quantity sensitivities.