A hybrid artificial neural networks approach to solve the inverse problem in advection-dispersion models

Abstract

In this paper, prediction capability of a hybrid Artificial Neural Networks (ANN) was investigated to solve the groundwater inverse problem. Initially, a Multi Layer Perceptron (MLP) network was developed and it was found that network produced better results when the target range of the parameters is smaller. Therefore, a Self-Organising Network (SON) was used to identify the objective subrange of the parameter and then the MLP model was employed to obtain final estimates. The data for the ANN was obtained from a numerical model that was utilised to simulate the solute transport in saturated groundwater flow. The forward problem of the numerical model was solved to generate solute concentration data for range of parameters. Those input data was fed into a MLP ANN to train the network along with corresponding parameter values. Sufficiently trained ANN model was used to estimate hydraulic conductivity (single parameter), and hydraulic conductivity and longitudinal dispersion coefficient (two parameters). First, the approach was tested on synthetic data to identify its feasibility and robustness. Then an experimental dataset that was obtained from an artificial aquifer was used to validate the method. It was found that ANNs produce accurate estimates in the presence of uncertainty. However, ANN are able to produce accurate results only if the pattern of the dataset that use to estimate parameters are similar to that of the training data. Therefore, it is important to adequately simulate the aquifer system in question by a large enough training dataset. However, due to the stochastic nature of the real world heterogeneous aquifers, it is not a trivial undertaking to identify the behaviour of the aquifer. Furthermore, as ANN's extrapolation capabilities beyond its calibration range is not reliable, it is necessary to set a calibration range sufficient to meet the limits of actual data. Therefore, prior information of the system is of utmost importance to obtain reasonably accurate estimates.