A cluster analysis approach to sampling domestic properties for sensor deployment.

Sensors are an increasingly widespread tool for monitoring utility usage (e.g., electricity) and environmental data (e.g., temperature). In large-scale projects, it is often impractical and sometimes impossible to place sensors at all sites of interest, for example due to limited sensor numbers or access. We test whether cluster analysis can be used to address this problem. We create clusters of potential sensor sites using factors that may influence sensor measurements. The clusters provide groups of sites that are similar to each other, and that differ between groups. Sampling a few sites from each group provides a subset that captures the diversity of sites. We test the approach with two types of sensors: utility usage (gas and water) and outdoor environment. Using a separate analysis for each sensor type, we create clusters using characteristics from up to 298 potential sites. We sample across these clusters to provide representative coverage for sensor installations. We verify the approach using data from the sensors installed as a result of the sampling, as well as using other sensor measures from all available sites over one year. Results show that sensor data vary across clusters, and vary with the factors used to create the clusters, thereby providing evidence that this cluster-based approach captures differences across sensor sites. This novel methodology provides representative sampling across potential sensor sites. It is generalisable to other sensor types and to any situation in which influencing factors at potential sites are known. We also discuss recommendations for future sensor-based large-scale projects. • A novel generalisable methodology for deployment of a limited number of sensors. • Cluster analysis allows data-driven unbiased sampling across possible sensor sites. • Clusters of similar homes are based on factors that influence the sensor measures. • Chosen sensor sites are verified using the resulting sensor data. • The methodology is applied and tested using gas, water and environmental sensors. [ABSTRACT FROM AUTHOR]