Daylighting performance prediction model for linear layouts of teaching building clusters utilizing deep learning

Deep learning (DL) has proven to be an effective tool for predicting the daylighting performance of buildings on individual rooms or standalone buildings by utilizing a few straightforward design parameters as input variables for analysis. In addition to existing studies, exploring methods to characterize larger objects with spatial relationships may contribute to understanding the impact of layout on the overall daylighting performance of buildings. In this study, a DL model based on the framework of “Autoencoder-Based Feature Extraction with Artificial Neural Network (AE-ANN)” has been developed to predict the daylighting performance of the layout of teaching building clusters. In order to efficiently extract the layout characteristics and improve the model's generalization capabilities, an autoencoder (AE) was pre-trained to encode the planar layout images of teaching building clusters into feature vectors, which were then employed for training an ANN model. In the testing dataset, the AE-ANN model demonstrated impressive accuracy, achieving R² values of 0.946 for sDA and 0.853 for ASE, alongside MSE values of 0.312 and 0.656. This research investigated the feasibility of the AE-based model for predicting daylighting performance of large-scale scenarios, highlighting its potential as a fundamental model for the development of more intricate daylighting prediction models.