Enhancing neighborhood-scale building performance simulation through building classification and automated data acquisition: Supporting the Dutch heating transition

Neighborhood-scale building performance simulation is essential for advancing the sustainable heating transition in the Netherlands. Such simulations require significant computing resources and are often hindered by a lack of access to detailed building information. This paper proposes a neighborhood-scale building performance simulation method enhanced with building classification and automated data acquisition to overcome these challenges. Deep learning-based models are developed to identify windows, doors, and photovoltaic (PV) panels from street view and satellite images. They enable the acquisition of critical modeling information that might be not readily accessible, including window/door area fractions, PV panel locations, and zoning configurations. Bayesian optimization is applied for model calibration to determine suitable values for uncertain building information (i.e., infiltration rates). Additionally, a clustering-based building classification approach is proposed to extract representative buildings from all the buildings in a neighborhood. Similar buildings usually have comparable heating performance, making it feasible to use a few representative buildings to simulate a large number of buildings. The proposed building performance simulation method is evaluated using the 1452 terraced houses located in a Dutch neighborhood. The identification accuracy of deep learning is 98.9 % for windows/doors and 98.6 % for PV panels. A total of 123 representative houses are extracted and modeled, leading to a 91.5 % reduction in simulation time. The representative house models exhibit a very small absolute percentage error (0.17 %) in simulating the neighborhood's annual gas consumption after model calibration.