Geospatial Graph Attention Network for High-Resolution Building Facade Photovoltaic Potential Prediction

Accurately predicting the photovoltaic (PV) potential of urban building facades plays a crucial role in the development of photovoltaics. This study proposes an innovative building facade PV potential prediction method based on the Geospatial Graph Attention Neural Network (GGAT). Compared to traditional methods, this approach considers the differences in solar radiation intensity at various heights of the building facade, enabling more precise identification of areas with higher PV potential on the facade. The study focuses on buildings in the Manhattan area of New York City and employs Rhino software and the Ladybug Tools plugin to conduct building solar radiation simulations, obtaining high-quality training data. During the modeling process, the concept of building height stratification is introduced, dividing the building facade vertically into 10 equal-height layers, with each prediction point representing the average solar radiation intensity within that height range. Experimental results indicate that GNN-based algorithms (especially GGAT) outperform traditional machine learning algorithms in predicting solar radiation on building facades. GGAT integrates geospatial features and graph attention mechanisms, enabling more accurate prediction of solar radiation on building facades. Solar radiation intensity exhibits significant differences both in the vertical direction of the building facade and in the horizontal direction (between census tracts). The stratified modeling method can reveal these differences, providing more comprehensive and detailed information for analyzing the PV potential of building facades.