Learning lighting models for optimal control of lighting system via experimental and numerical approach

Lighting control systems have been traditionally employed to reduce energy use for lighting by, for example, maximizing daylight harvesting. When highly efficient light sources are installed and for tasks where maintaining target illuminance is particularly important, designers may decide to prioritize the latter together with energy use. In this context, the use of data-driven algorithms is emerging. In this paper different data-driven approaches are proposed as lighting control systems, to maximize daylight harvesting and to optimize energy consumption. The approaches employ experimental data of occupancy and lighting switch on/off events of a private side-lit office in an academic building. The office is later modeled in DIVA4Rhino to provide yearly illuminances and electric lighting dimming profiles. These data are used to implement data-driven optimal controls. Three different approaches have been employed: Regression Trees; Random Forests; Least Squares. Different lighting control strategies have been hypothesized based on installed Lighting Power Densities (LPD). Results show that Regression Trees outperforms both Least Squares and Random Forests, in terms of model accuracy and control performance.