Using support vector machine to detect desk illuminance sensor blockage for closed-loop daylight harvesting
Michael Kent, Nam Khoa Huynh, Stefano Schiavon, Stephen Selkowitz
Abstract
Daylight can reduce electric lighting in buildings. This is facilitated by sensors that relay real-time illuminance data to a light controller. When daylight provides greater than required or desired levels of illuminance, control actions enable electric lights to reduce their output and save energy. Occupant behaviours can block desk sensors and this reduces the amount of energy saved. However, no method exists that can be used to continuously monitor sensors to ensure they operate as intended (e.g. remain unblocked). We carried out a study in an open-plan office building in Singapore, consisting of 39 workstations each fitted with desk illuminance sensors independently controlling a dedicated ceiling light. Power over Ethernet was used to collect individual data signals for both illuminance and power from each workstation. Data were collected across a one month period, sampling signals at every 2-minute interval. A linear support vector machine model accurately classified 99% of the data points using our sensor blocking algorithm. From 447,455 data points analysed, 12% of dataset showed that sensors were blocked and this had an estimated energy penalty of 24%. We do not recommend installing illuminance sensors at the desk. Our study highlights the usefulness of Power over Ethernet for closed-loop daylight harvesting. The data collected can be used to monitor the health of the sensors’ performance to help minimise energy use.