Daylight Prediction Surrogate Using Radiance, Honeybee and Machine Learning

How much daylight will a classroom or studio get - before the drawings are finished? A Radiance simulation answers this accurately but needs seconds to minutes per design; the classical BRE average daylight-factor formula is instant but ignores room depth, overhangs and the distribution of light. This project builds its own Radiance data set of 400 shoebox rooms (Latin hypercube over ten design variables, CIE overcast sky, daylight factor on a 0.5 m work-plane grid), cross-checks it with Ladybug Tools / Honeybee and finer Radiance settings, and trains machine- learning surrogates - gradient-boosted trees, a Gaussian process with an uncertainty band, a physics-informed GP on the BRE ratio and a PCA + GP model of the whole daylight-factor map - that answer in milliseconds. They are compared honestly with the BRE formula and linear regression, tested on rooms deeper than any training room, and used for a teaching compliance check and a quick design exploration whose answers are re-simulated in Radiance. A web application on five databases checks rooms, draws the predicted DF map and verifies any answer with Radiance.

Category: School of Architecture and Planning

Architecture & Planning Project AP03 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.