Key control variables affecting interior visual comfort for automated louver control in open-plan office -- a study using machine learning

Model-based control strategy has been a promising approach for maximizing the potentials of automated louver systems regarding visual comfort optimizations. As the data-driven approach is thriving, the machine learning algorithm can enhance the efficiency of model-based shading control by building a statically predictive daylighting model. Determining daylighting-affected variables is a crucial step benefiting approximation accuracy and modeling efficiency. However, studies about variable selections are rare since it is intractable and expertise required, and those limited studies mainly focus on private offices as a research subject. This paper performs a comprehensive analysis to explore the importance of control-related variables available for machine learning-assisted automated louvers in open-plan offices. The aim is to provide an efficient but compact set of control variables for better establishing the predictive model regarding visual comfort assisting automatic shading control for optimizing the daylighting environment. First, a validated simulation was performed to generate data samples that were used for feature analysis. Then, Filters, Embedded, and Wrappers methods were imposed to interpret feature importance according to their contributions to the robustness of the predictive model. Multicollinearity analysis was then employed to eliminate unnecessary features further. Finally, attributes affecting visual comfort in an open-plan office were derived from a comprehensive comparison and validated via a model training test. Results indicated that spatiotemporal occupancy info weighs most in affecting the predictive model, and the selected variables were proved efficient in the prediction horizon test.