Using WiFi connection counts and camera-based occupancy counts to estimate and predict building occupancy: Fid-Bau Portal

Using WiFi connection counts and camera-based occupancy counts to estimate and predict building occupancy

Alishahi, Nastaran / Ouf, Mohamed M. / Nik-Bakht, Mazdak

Highlights • Long-term ground truth occupancy data was collected via camera-based image recognition counters. • WiFi-Occupancy conversion factor was highly affected by temporal attributes and occupancy level. • WiFi-Occupancy conversion factor tended to converge to a fixed threshold during peak hours. • Day-ahead occupancy was predicted with high accuracy using WiFi counts of the previous day. • Importance of considering temporal and occupancy variation when using WiFi counts was highlighted.

Abstract Accurate occupancy information can help in optimizing the operation of building systems. To obtain this information, previous studies suggested using WiFi connection counts due to their strong correlation with occupancy counts. However, validating this correlation and investigating its variation have remained limited due to challenges regarding the collection of ground-truth data. Therefore, many studies suggested a single (fixed) value as the conversion factor of WiFi connection counts to actual occupancy counts based on short-term ground-truth data. This study addressed this gap by proposing a method for investigating the correlation between WiFi connection counts and actual building occupancy over a longer duration using continuous ground-truth data collected from camera-based occupancy counters. The proposed method focused on (i) identifying the influential features on this correlation and their effectiveness, as well as (ii) developing models to estimate real-time occupancy counts and to forecast day-ahead occupancy counts. To validate the proposed method, it was applied in a library building in Montréal, Canada with data collected between January and March 2020. Results showed time-related features including Hour of the day and Day of the week, as well as occupancy level influenced the correlation between Wifi and occupancy counts. Furthermore, the proposed models successfully estimated real-time occupancy counts with an average accuracy (R²) of 0.96 for weekdays and 0.98 for weekends, while day-ahead occupancy forecasting models had an average accuracy (R²) of 0.92 for weekdays and 0.82 for weekends. These findings provided a proof-of-concept for the proposed methodology, demonstrated the potential of using WiFi connection count for estimating/forecasting occupancy counts, and most importantly highlighted the important considerations that need to be addressed when using WiFi connection counts as a proxy for occupancy to optimize building operation.