Developing a parametric morphable annual daylight prediction model with improved generalization capability for the early stages of office building design

Building form and fenestration design decisions made in the early stages of design have considerable impact on the annual daylight performance of office buildings. Annual daylight performance needs to be evaluated at the conceptual design stage to support the building form and fenestration design decision-making process. However, the simulation modeling and ray-trace calculation required for annual daylight prediction are extremely time consuming, with an adverse impact on its feasibility in the early design stages. Machine learning-based models have received much attention to reduce the daylight simulation time; however, the generalization capability of these models is limited. This study develops an artificial neural network-based modeling approach to predict annual daylight performance in the early stages of the design process. A workflow to develop an annual daylight prediction model with higher generalization capability is proposed, through feature selection, feature engineering, and hyperparameter optimization, with an accompanying tool to integrate the machine learning model into the early design environment. The developed prediction model was validated against Radiance simulation results with a high accuracy setting and attained R² scores of 0.988 and 0.996, MAE scores of 1.58 and 1.37, MAPE scores of 2.10% and 2.36% for UDI and DA₃₀₀, respectively, while being 250 times faster. The proposed modeling approach can be extended by adding more types of parametric room modules.