Overcoming Computational Limits in Classroom Lighting Design through Machine Learning–Based Prediction

Improving daylighting conditions typically conflicts with energy savings goals. Multiobjective optimization methods are often used to balance these competing design aspects, but their high computational time, especially for daylight simulations, limits their practical application. In this study, machine learning (ML) models were used to predict daylighting and energy performance in US classrooms among four climate zones. A leave-one-climate-out cross-validation scheme was used to test the generalizability to other climates. The parameters examined include classroom dimensions, window head height, orientation, window-to-wall ratio (WWR), and window width in single- and double-loaded school typologies. Results show that Random Forest repeatedly provided the highest prediction accuracy for daylighting performance, while Support Vector Machine and Artificial Neural Network performed best for energy performance. The analysis shows that for the single-loaded corridor typology, orientation, and WWR are the most influential factors across all optimization objectives. For the double-loaded corridor typology, classroom geometry, expressed by the width-to-depth ratio, is the dominant parameter affecting total energy use. Daylighting performance at the horizontal desk-plane is mainly affected by WWR, orientation, and width-to-depth ratio, whereas daylighting at the vertical eye-plane varies among climate zones. These findings demonstrate that ML models can reliably predict daylighting and energy outcomes and highlight the key design drivers for optimizing classroom performance across US climate zones. The results can support early design decision-making and reduce the computational burden of simulation-based optimization.