Data Collection Period and Sensor Selection Method for Smart Building Occupancy Prediction

Building energy consumption depends on many factors, such as occupant behavior and occupancy. Many works study building occupancy modeling and its impact on energy consumption, based on sensors, such as CO2, humidity, presence, etc., which are deployed within buildings and their surrounding areas. These sensors collect different types of data at a high frequency, which can be used to build datasets used in building predictive models. In this context, existing datasets have been empirically built without considering the relevant sensor types and the frequency of data collection for building occupancy modeling. Therefore, in this paper, we introduce a method to select the data collection period and the relevant sensors for building occupancy prediction model with satisfying accuracy. Our approach uses feature selection and machine learning classifier algorithms, which are applied to different data collection periods, starting from 1 minute to 60 minutes. The experiments, carried out on a real dataset, with 5 different machine learning classifiers show that it is possible to build an occupancy predictive model with Random Forest having an accuracy of at least 90%, by using 8 sensors collecting data at a 20-min interval, or 5 sensors collecting data at a 15-min interval.