A parametric design integrated sampling and general training approach for optimal control oriented surrogate models of light-related quantities

Abstract

This study presents a general method for determining optimal control oriented surrogate models of light-related quantities. It is termed the incidence operator method comprising of sampling, model training and model export. In contrast to matrix-based methods, machine learning constitutes a fundamental component of this new method. This entails the ability to represent complex dependencies of light-related quantities on adjustable material and geometric properties, as well as the possibility to export models using a standardised format (e.g. FMI, ONNX). Furthermore, components were developed to streamline the sampling of training data from parametric designs. The associated higher resolution reveals spatial discontinuities, for which a novel modelling approach and integration methodology have been developed. The incidence operator method was validated with the enhanced two-phase method using two scenes. Based on annual simulations, the simple scene demonstrates high agreement (nMAE<2.1%), while the complex fenestration scene exhibits spatial discontinuities resulting in an increased deviation (nMAE<11.9%).