Non-intrusive hourly occupancy detection in residential buildings using remotely readable water meter data: Validation and large-scale analysis

Occupancy significantly influences building energy use; however, large-scale occupancy estimation, particularly in residential buildings, remains a challenge. While surveys and sensor data, particularly from remotely readable electricity meters, have been used for non-intrusive occupancy detection, water meter data have so far only been applied at daily resolution. This study introduces, validates, and applies a novel algorithm to infer hourly occupancy using data from remotely readable water meters. A new day definition, spanning noon to noon, is proposed to reflect typical daily rhythms better. Daily occupancy is inferred using water usage events and average consumption, akin to established methods. For days classified as occupied, hourly occupancy is estimated by associating water use with occupancy, followed by smoothing. Nighttime occupancy is inferred by linking the last evening and first morning occupancy. Validation was performed using one year of manually labelled data from ten households for daily occupancy (MCC = 0.982), and 147 days of unseen ground-truth data from nine households for hourly occupancy (MCC range: 0.267 to 0.795; building-weighted MCC = 0.594). Short absences were often misclassified as occupied, while longer absences ( > 3 h) were reliably detected. Comparison with a state-of-the-art electricity-based algorithm showed superior performance of the proposed water-based method. Finally, the algorithm was applied to one year of data from 2,690 households, yielding plausible and interpretable occupancy patterns at both daily and hourly scales. These results demonstrate the method’s robustness and scalability, offering a promising approach for large-scale, non-intrusive occupancy monitoring. • Non-intrusive occupant monitoring to obtain hourly occupancy from hourly water data. • Coherent occupancy on daily and hourly timescales for residential buildings. • Overall occupancy can be inferred with high accuracy, but variations across buildings. • Proposed algorithm is superior to an electricity-based algorithm from the literature. • Algorithm applied to 2690 houses shows valid, scalable, and explainable results.