Window extraction from aerial photogrammetry point cloud datasets for the development of energy digital twins (EDTs)

Accurate geometric extraction of building envelope elements from 3D point clouds is fundamental to developing Energy Digital Twins (EDTs) grounded in geospatial datasets. The precise extraction of windows from aerial photogrammetry is essential for the simulation of building energy performance as they have a significant impact on solar gains, heat loss, and daylighting. Using aerial photogrammetry point cloud dataset from metropolitan area in Torino, Italy, this study compares two methods for automatic window extraction: a Random Forest (RF) classifier trained on manually defined geometric features and a Kernel Point Convolution (KPConv) network that captures hierarchical geometric features from unstructured point clouds. While the RF model attained an overall accuracy of 56.2% with a window-class F1-score of 38.3%, KPConv exhibited improved performance with a 52.1% F1-score and 66.4% accuracy, indicating its relatively greater reliability in capturing window geometry. In software like EnergyPlus, the correct computation of energy performance is enabled by the detailed depiction of building attributes such as windows, which aids in simulating heating, cooling, and lighting demands. Inaccurate assessment of solar gains and thermal losses may arise from inaccuracies in window shape, thereby affecting energy demand forecasts. These findings underscore the significance of superior geometric extraction in the formulation of efficient EDT, as it creates a scalable framework for energy-efficient architectural design, retrofitting, and sustainable urban planning on a large scale. This work illustrates the extraction of windows from aerial photogrammetry point cloud datasets as a fundamental step in the development of Energy Digital Twins (EDTs), supplying crucial boundary inputs for later building energy modelling.