Integrating multimodal data using machine learning methods to identify prototype building stocks for urban building energy consumption assessment

Assessing the energy consumption of urban buildings is critical for managing city carbon emissions. Given the large number of urban buildings, it is essential to identify the stock of urban prototype buildings. However, current research faces challenges in linking building feature identification with energy consumption evaluation, highlighting the need for methods that improve the functional and spatial coverage of building identification. We developed a comprehensive dataset of urban building footprints and proposed a hybrid approach combining poin of interest identification and machine learning to classify the functions of urban buildings. In addition, we identified the construction year and scale of buildings, analyzing the stock and spatial distribution of 15 types of prototype buildings. The results show that our hybrid method performs well in identifying 22 different building functions, with an overall classification accuracy of 0.86 for the machine learning component. The analysis of prototype building stock reveals that residential buildings are primarily low-rise and multi-story, while public buildings are dominated by offices and schools. Buildings from different periods exhibit distinct spatial distribution patterns. Our study provides a data foundation for evaluating building energy consumption at the urban scale, supporting sustainable urban management efforts. © 2025 Elsevier Ltd