Occupancy prediction using deep learning approaches across multiple space types: A minimum sensing strategy

The proliferation of sensing technologies has allowed the collection of occupancy-related data to support various building applications, including adaptive HVAC and lighting controls, maintenance operations, and space utilisation. However, past occupancy prediction studies often considered different combinations of sensor data and investigated a limited number of space types. This study performs occupancy prediction based on a minimum sensing strategy by using a comprehensive set of sensor data (i.e., indoor environmental and outdoor weather conditions, Wi-Fi connected devices, energy consumption data, HVAC operations, and time-related information) to identify the most crucial features through a proposed feature selection algorithm. Occupancy predictions were subsequently performed using different deep learning architectures, including Deep Neural Network (DNN), Long Short-Term Memory (LSTM), Bi-directional LSTM (Bi-LSTM), Gated Recurrent Unit (GRU), and Bi-directional GRU (Bi-GRU) in an office, library, and lecture room. Our findings highlighted that the proposed feature selection algorithm outperformed a popular feature selection algorithm to achieve a higher model performance with lower sensing requirements. Furthermore, empirical results showed that indoor CO2 levels and Wi-Fi connected devices were crucial features for predicting occupancy across all space types. The best model performances were achieved using Bi-GRU for office, GRU for library, and Bi-GRU for lecture room.

Automated summary

This study performs occupancy prediction based on a minimum sensing strategy using a comprehensive set of sensor data and a proposed feature selection algorithm. The findings highlight the crucial features for predicting occupancy across all space types as indoor CO2 levels and Wi-Fi connected devices.