Deep Learning Techniques for Prediction of Non-visual Luminous Content of Cellular Offices
Abstract
Simulation evaluation of non-image-forming (NIF) effects of daylight in the built environment necessitates using computationally demanding and specialised software. Therefore, this study introduces an alternative approach by exploring the potential of implementing Artificial Neural Networks (ANNs) to predict NIF effects in unilaterally daylit rectangular office spaces. The ANN models were trained on a dataset generated by simulating 349,445 cases of various office geometric configurations, optical material properties, location, sky types, and time of day in a year. The Circadian Stimulus model achieved the best ANN regression model performance with R2 of 0.965, while the melanopic Equivalent Daylight Illuminance model predicted compliance with minimum requirements with 96.7 % accuracy. The results show the practical implications of ANN models for fast prediction of NIF effects in the built environment, significantly reducing the time and effort required for such evaluations and particularly suited for early-stage design phases.