Virtual sensors to improve on-line hydraulic model calibration

A new approach for monitoring a water distribution system involving Virtual Sensors is presented. In this approach, wireless sensor nodes are permanently deployed within the distribution system, providing continuous, on-line hydraulic data that can be assimilated into hydraulic models. In addition, temporary nodes are deployed for short periods (one week) around the distribution network. A Virtual Sensor is implemented using a data imputation technique called Gaussian Process Regression, which combines the historical data collected by the temporary node with correlated data from a subset of permanent sensor nodes. Use of spatially-correlated data accounts for new trends in the data that do not appear in the historical data collected by the temporary node. An increase in the number of sensors (a combination of real and virtual) is important for reducing the ill-conditioned state of the hydraulic model calibration procedure. The technique is demonstrated as a proof-of-concept using data collected from the WaterWiSe@SG testbed in Singapore, and is shown to predict pressure data trends with an accuracy of 0.76 PSI RMSE after a six-week test.