Daylight factor prediction using machine learning: A two-way study using numerical encoding and regression models, versus image encoding and pix2pix

This paper proposes a Machine-learning-based methodology for the real-time prediction of Daylight Factor (DF) during the conceptual design phase of architecture. The predictions allow the designer to quickly assess DF regulatory compliance and support decision-making at early design stages. To achieve this goal, a digital surrogate dataset is generated from a real-life design scenario. Image-based and numerical-based modelling approaches are developed, deployed, evaluated, and reported in this paper, supported by different encodings derived from the dataset. The first method uses point-by-point DF prediction through numerical geometrical and performance data encoding. It trains numerical regression models, namely an Artificial Neural Network and XGBoost model. The second approach uses image-based encoding. The dataset is translated into false colour images of the DF values per room floor plan with a 128 x 128 resolution and is used to train a Pix2Pix model. The paper compares the two modelling approaches and evaluates their advantages and disadvantages, demonstrated through the early-stage design use case. Results have shown that both approaches can accurately interpret DF prediction quantitatively with sufficient accuracy; however, the qualitative implications of the two modelling approaches are discussed.