Predictive models for daylight performance of general floorplans based on CNN and GAN: A proof-of-concept study

Abstract

A daylight performance evaluation at the early design stage is essential for a building morphology design and optimization, having a tremendous influence on energy consumption and indoor environments. Considering the complicated input parameters, time and computational cost of the simulation tools, although proxy models based on various machine learning algorithms have been developed, they are limited to certain building forms described based on the selected parameters. In this study, proxy models of a daylight simulation for general floorplans are proposed based on convolutional neural network (CNN) and generative adversarial network (GAN). ResNet (CNN) and pix2pix (GAN) are applied to predict static and annual daylight metrics (uniformity, mean lux, success rate, sDA, and UDI) and illuminance distribution in space, respectively. Geometry information is embedded in the image structure, and a grid-based simulation are conducted as the ground truth. Two datasets composed of real floorplan cases and parametric rooms were tested in the experiments. ResNet obtains the best R2 of 0.959 and an MSE of 0.008 on the real case dataset for daylight uniformity. In addition, pix2pix generates visualization results close to the simulation with an SSIM of 0.90 in the test set within a period of 1 s and provides real-time intuitive feedback for designers. The results show the possibility of using deep neural networks to extract features from general building forms and build predictive models, which can be integrated into automatic form-finding and design optimization.