Using machine learning algorithm to predict lighting energy consumption of daylight-linked lighting systems from spatial daylight autonomy

Abstract

Assessing the energy savings from daylight linked control (DLC) for lighting systems in a generalized form is challenging. However, novel approaches, such as the Machine learning algorithm (MLA), have the potential to address this challenge and are worth further investigation. This study aims to predict the energy consumption of lighting systems with DLC from the daylighting performance metric: spatial daylight autonomy (sDA), along with several necessary design features. A parametric room model with single side-lit window is established, and four DLC modes are set. From these, around sixteen thousand data sets comprising sDA, room design features and lighting energy consumption are collected for training the algorithm model. The XGBoost model is selected as it outperforms other algorithms by accuracy and efficiency. The results of data analysis demonstrate that the prediction model developed with sDA and several design features exhibits commendable predictive performance, and these features include Room Area, Room Length, Room Height, WWR, and Room Width. The following conclusions can be drawn: sDA along with four to six design features, depending on control mode, are effective for predicting the energy consumption of a lighting system applying DLC in rooms of varied dimensions. The XGBoost has demonstrated efficacy in addressing regression issues and managing the complex nonlinear relationships inherent in dynamic daylighting related issues. The model produced decisive data and provided a rapid method that assists decision-makers in choosing between DLC and conventional lighting control systems. It is also a meaningful exploration of AI applications for building daylighting performance analysis.