A graph-based hybrid deep learning approach for the thermal performance potential prediction of green roofs

Green roofs have emerged as passive energy-saving technologies in building energy efficiency design and retrofit. In order to bolster the role of green roofs in building energy efficiency, the thermal performance of green roofs needs to be captured. In this paper, a novel artificial intelligence (AI) model framework integrating graph convolution is presented for predicting interior and exterior surface temperatures of the green roof. The framework incorporates the graph convolutional network (GCN), the fully connected network (FCN), and the SHapley Additive exPlanations (SHAP). The model training dataset is established using meteorological data collected in Shanghai, China and experimental data from green roofs outside the existing building. The model's predictive performance is evaluated by various assessment metrics. The results show that the model has high prediction accuracy for the two tasks, with average R-square (R-2), mean absolute error (MAE), and root mean square error (RMSE) of 0.979, 0.326 C-degrees, and 0.410 C-degrees, respectively. The SHAP results reveal that the soil surface temperature and the canopy temperature are the main factors affecting the thermal performance of green roofs. The prediction accuracy of our proposed model is better than other similar models, which provides an effective guide on the application of green roofs in building energy efficiency retrofits and implementations.