Deep Semi-Supervised Multi-Task Learning of Building Features for District Energy Demand Estimation

Building energy demand is a primary driver of greenhouse gas emissions, necessitating accurate, sector-coupled energy system analysis via co-simulation. However, parameterizing white-box building models is frequently hindered by the unavailability of essential features in public databases. We present a method to estimate missing attributes—specifically construction year (CY), building type (BT), and energy carrier (EC)—using street-level imagery (SLI) across Germany as an exemplary use case. Our automated workflow integrates SLI with 2022 German Census data for labeling and is validated against OpenStreetMap (OSM) building geometries. A single deep learning model is developed using multi-task learning (MTL) and combined with semi-supervised learning (SSL) to effectively leverage partially labeled datasets. While the results demonstrate strong generalization to unseen test data, performance remains constrained by data quality issues that impact full automation.