Generating a nationwide residential building types dataset using machine learning

• Assignment of TABULA archetypes to German residential buildings. • XGBoost model suitable for size class and construction year imputation. • Census data serves as input for training data generation. • Features from 3-D building data, OpenStreetMap, census, and socioeconomic datasets. • Openly available complete nation-wide building archetype dataset for Germany. The lack of high-resolution building data is an obstacle to the development of detailed, spatially explicit recommendations for decarbonization measures. In an effort to fill this gap, this study outlines the creation of a building level dataset based on standardized building archetypes for all German residential buildings. A machine learning approach using XGBoost is used to train models to predict the size class and construction year of individual buildings. Refurbishment states are assigned based on federal state level statistics. Based on these characteristics, TABULA building archetypes are assigned. The training data generation is primarily based on the grid dataset of the German census. The data is enriched with morphological features of buildings and neighborhoods, as well as socio-economic characteristics. The machine learning models perform with accuracies of 97.4 % and 73.9 %, respectively, on a test set at the individual building level. The distribution of size classes and construction years in the resulting dataset shows a high degree of agreement with official statistics at the federal state level, but also a tendency to overrepresent majority classes. This study proves that the chosen methodology is suitable for generating a complete nationwide dataset. By providing spatially resolved, individual building data that can serve as a proxy for the energetic properties of buildings, the resulting dataset can facilitate building-related energy transition analyses.