There is a growing desire to measure the operational performance of buildings – often many buildings simultaneously – but the cost of sensors and complexity of deployment is a significant constraint. In this paper, we present an approach to minimising the cost of sensing by recognising that researchers are often not interested in the raw data itself but rather some inferred performance metric (e.g. high CO2 levels may indicate poor ventilation). We cast the problem as one of constrained optimisation – specifically, as a bounded knapsack problem (BKP) – to choose the best sensors for the set given each sensor's predictive value and cost. Training data is obtained from a field study comprising a wide range of possible sensors from which a minimum set can be extracted. We validate the method using reliable self-reported event diaries as a measure of actual performance. Results show that the method produces sensors sets that are good predictors of performance and the optimal sets vary substantially with the constraint parameters. Furthermore, valuable yet expensive sensors are often not chosen in the optimal set due to strong co-incidence of sensor signals. For example, light level and sound level often increase at the same time. The overall implication of the work is that a large number of co-incident low-cost sensors can be used to build up a picture of building performance, without significantly compromising information content, and this could have major benefits for the smart metering industry.
ENLITEN - A dataset and its associated analysis code for the paper entitled "Designing sensor sets for capturing energy events in buildings"