Developing an Advanced Daylight Model for a Shading System with Conditional Generative Adversarial Network (cGAN)

This paper presents a conditional Generative Adversarial Network (cGAN)-based image-to-image translation method to predict indoor illuminance with a complex shading system. Using illuminance as the daylight metric is computationally expensive, due to the time-consuming ray trace and/or radiosity methods required to conduct multiple simulations. The major contribution of the research lies in the proposed model that predicts indoor illuminance in less than a second while maintaining an 84% prediction accuracy. The predicted outcome is in the form of a contour plot, providing intuitive feedback to designers during the early design phase, unlike traditional models which predict a numeric value. The contour plot can potentially be filtered by color to extract the minimum and maximum illuminances during different times, allowing designers to identify periods with high contrast that create undesirable daylight distribution. The instantaneous daylight prediction can later be utilized to design and optimize a complex shading system.