Identification of Influential User Locations for Smart Meter Installation to Reconstruct the Urban Demand Pattern

This paper aims to show how the total demand signal (i.e., the sum of all user demands) of a district metered area (DMA) can be reconstructed using simple linear models starting from the readings taken at influential user locations with a preassigned temporal resolution (e.g., 1 h). By comparing the total reconstructed demand pattern with the inflow into the DMA, which is usually monitored at DMA boundaries using flowmeters, water utilities have the potential to identify occurrences of anomalous events, such as pipe bursts, unauthorized consumption, and leakage increases. To address this issue, two procedures are proposed that are applicable when smart meters at the end of life were previously installed at all locations and when no smart meter was present, respectively. The first procedure is based on the application of the stepwise regression that measures the demand time series aimed at both the selection of user locations and the model calibration, whereas Fisher’s test or economic considerations is used as stopping criterion. The second methodology consists of the application of criteria to identify the subset of representative locations on the basis of available data and practical considerations (user typology and consumption on the annual bill). In addition, a new method to calibrate the model using billed annual consumptions is provided. Both methodologies enable the construction of linear models that express the total pattern as a function of single-user consumption patterns, allowing the reconstruction of the original signal. Both procedures were applied to a case study in Naples (Italy) with a total of 1,406 users, that is, 1,067 residential users and 339 non-residential users. The results proved that, as expected, the accuracy of the total demand pattern reconstruction of both procedures increases as the sample size of the representative locations grows. However, the results indicating the trade-off between the sample size and the goodness of the fit reveal that the accuracy is good even for low model order (25–50 users selected). Indeed, the coefficient of determination, $R^2$, of the fit increases from 0.66 to 0.83 for 25 and 50 users selected, respectively, and the accuracy becomes almost perfect for a size of 100 (i.e., approximately 7% of the total number of users considered). Overall, this paper demonstrates that an effective strategy for the accurate characterization of the total demand in a DMA consists of distributing the preassigned number of smart meters between the different categories of users present in that DMA (i.e., residential and non-residential) and privileging users with the highest annual consumption.