Network design for predicting groundwater contamination

Abstract

A new network design algorithm for improving the reliability of groundwater simulation model predictions is developed. The objective of the algorithm is to minimize the simulation model prediction variance by choice of new aquifer property measurement locations. This method, which uses parameter measurements to minimize prediction error, is different in concept from conventional parameter estimation or inverse methods which use state variable measurements to minimize parameter estimation error. The variance of predicted state variables, hydraulic head and contaminant concentration, is used as a measure of model prediction reliability. Prediction variance depends on the extent and quality of aquifer property measurements used to estimate simulation model parameters. Kriging is used in the first‐order uncertainty analysis to estimate the model parameters. First‐order uncertainty analysis is used in the algorithm to compute the prediction variance and select new measurement locations. A network design example is presented, the results of which show that a significant increase in simulation model prediction reliability is achieved by measuring aquifer properties at locations selected by the algorithm. The predicted mean was found to be insensitive to new measurements, while the predicted state variable variance was found to be quite sensitive to new measurements. The selection of measurement locations was found to be greatly affected by the type and extent of boundary conditions existing in the aquifer. Fixed value boundary conditions tend to reduce prediction variance in the neighborhood of new measurements and throughout the entire aquifer. Specified flux and mixed boundary conditions tend to reduce prediction variance in the neighborhood of the measurement locations, with little reduction occurring in other areas of the aquifer. The results also indicate that both the mean head and concentration are relatively insensitive to the number of data points in the network as long as there is a good estimate of the statistics of the ln (K) distribution and covariance function. This is not the case with the standard deviation of both the head and concentration which are significantly affected by the addition of data points to the network.