Virtual PMV sensor towards smart thermostats: Comparison of modeling approaches using intrusive data

Smart thermostats are considered an effective digital technology for building energy efficiency. This study proposes a virtual predicted mean vote (PMV) sensor for smart thermostats for indoor thermal comfort and energy-efficient building operation and control. The proposed virtual sensor can observe the representative PMV for the occupied zone, considering the systematic errors caused by the spatial difference between the wallmounted thermostat and the occupied zone. Virtual PMV sensor-based smart thermostat has the potential to improve energy efficiency and thermal comfort by observing unmeasured thermal comfort within the occupied zone based on variables obtained from an existing wall-installed thermostat. To cover systematic errors for general applications, intrusive measurements using a commissioning period are necessary, even in the short term. Thus, conventional modeling approaches can be used with limited intrusive data to develop an accurate insitu virtual PMV sensor during operation. In the application into a small-sized office building, the real systematic errors showed up to -2.4 degrees C among ten different thermostat locations in the target office zone. The in-situ virtual sensor could accurately observe the PMV for an occupied zone for different air-conditioning periods (airconditioned working hours, non-air-conditioned working hours, and closing hours) with the best accuracy of 0.05 mean absolute error (MAE) using one-day intrusive dataset. Based on comparing white-box, black-box, and gray-box virtual sensors in the application, this study recommends using the gray-box modeling approach to leverage the synergetic effect between data and knowledge for better in-situ virtual sensor modeling under a limited intrusive dataset.

Automated summary

This study proposes a virtual predicted mean vote (PMV) sensor for smart thermostats to improve energy efficiency and thermal comfort. The study recommends using the gray-box modeling approach to leverage the synergetic effect between data and knowledge for better in-situ virtual sensor modeling.