Opportunistic occupancy-count estimation using sensor fusion: A case study

• Published in: Building and environment 2019-07, Vol.159, p.106154, Article 106154
• Main Authors: Hobson, Brodie W., Lowcay, Daniel, Gunay, H. Burak, Ashouri, Araz, Newsham, Guy R.
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Buildings; Carbon dioxide; Case studies; Commercial buildings; Cost analysis; Data transmission; Electrical loads; Energy management; Formalism; Ground truth; HVAC equipment; Measuring instruments; Motion detection; Neural networks; Occupancy; Occupancy detection; Occupancy-centric controls; Occupancy-count estimation; Office buildings; Regression analysis; Regression models; Sensor fusion
• ISSN: 0360-1323; 1873-684X
• DOI: 10.1016/j.buildenv.2019.05.032

Estimation of occupancy counts in commercial and institutional buildings enables enhanced energy-use management and workspace allocation. This paper presents the analysis of cost-effective, opportunistic data streams from an academic office building to develop occupancy-count estimations for HVAC control purposes. Implicit occupancy sensing via sensor fusion is conducted using available data from Wi-Fi access points, CO2 sensors, PIR motion detectors, and plug and light electricity load meters, with over 200 h of concurrent ground truth occupancy counts. Multiple linear regression and artificial neural network model formalisms are employed to blend these individual data streams in an exhaustive number of combinations. The findings suggest that multiple linear regression models are the superior model formalism when model transferability between floors is of high value in the case study building. Wi-Fi enabled device counts are shown to have high utility for occupancy-count estimations with a mean R2 of 80.1–83.0% compared to ground truth counts during occupied hours. Aggregated electrical load data are shown to be of higher utility than separately submetered plug and lighting load data. •PIR, CO2, lighting and plug loads, and Wi-Fi data were collected from an academic building.•Models for occupancy-count estimation were developed with concurrent groundtruth data.•Wi-Fi enabled device counts were the best regressor to characterize occupancy levels.•Use of ANNs instead of MLR models reduced the transferability of occupancy-count models.•Dataset was made publicly available to support occupancy model development and assessment.