Occupancy count, i.e., the number of people in a space or building, is becoming an increasingly important measurement to model, predict, and minimize operational energy consumption. Explicit, hardware-based, occupancy counters have been proposed but wide scale adoption is limited due to the cost and invasiveness of system implementation. As an alternative approach, researchers propose using data from existing information and communication technology (ICT) systems to infer occupancy counts.

In the reported work, three different data streams, security access data, wireless connectivity data, and computer activity data, from ICT systems in a medium sized office building were collected and compared to the counts of a commercially available occupancy counter over 59 working days. The occupancy counts from the ICT systems are compared to the commercial counter with and without calibration to determine the ability of the data sets to measure occupancy. Various transformations were explored as calibration techniques for the ICT data sets. Training sets of 24, 48, and 120 hours were employed to determine how long an external calibration system would need to be installed.

The analysis found that calibration is required to provide accurate counts. While each ICT data set provides similar magnitudes and time series behavior, incorporating all three data streams in a two layer neural network with 1 week of training data provides the most accurate estimates against 5 performance metrics. Whilst 1 week of data provides the best results, 24 hours is sufficient to develop similar levels of performance.