Contaminant spread forecasting and confirmatory sampling location identification in a water-distribution system

While significant emphasis has been placed on contamination warning system design, event detection, and source identification, relatively little emphasis has been placed on characterizing contaminant spread and confirmatory sampling. This study developed algorithms that utilize contamination source probabilities to forecast contaminant spread, which is then used to identify confirmatory sampling locations to maximize contaminant spread information based on entropy concepts. The algorithms were applied to simulated contamination scenarios using one small network with five-sensor locations, and one large network with 5-, 10-, 20-, or 50-sensor locations. The first step in the forecasting process was to identify the past contamination probability using existing sensor information and a probabilistic contamination source identification algorithm. In general, the past contamination status of the nodes was either correctly identified or did not have enough information to classify; incorrect classifications were typically less than 3%. The past contamination status was then used to forecast contaminant spread 4 h into the future with the classification performance for the correct (32-53%) and incorrect (6-17%) identification, and unable to classify (30-57%) categories directly dependent on the accuracy of past contamination characterization. In general, forecasting accuracy increased with the number of sensor locations and decreased with longer forecasting time horizons. Confirmatory sampling locations were identified using the forecasted spread information and compared with locations selected using actual (i.e., perfectly known) spread for comparison. Both the forecasted and actual approaches generated similar confirmatory sampling locations. In general, the confirmatory sampling locations identified with the estimated forecasts created new information (i.e., reduced the number of nodes with unknown status), while using the actual forecasts reinforced existing information. As a result of the new information generated, the confirmatory sampling locations identified with the estimated forecasts tended to better characterize the contamination event than the locations identified with the actual sensor information known up to the future time of sampling.