A hybrid neural networks and numerical models approach for predicting groundwater abstraction impacts S. J. Birkinshaw, G. Parkin and Z. Rao ABSTRACT S. J. Birkinshaw (corresponding author) G. Parkin School of Civil Engineering and Geosciences, University of Newcastle upon Tyne, Newcastle NE1 7RU, UK Tel.: +44 191 2226 319 E-mail: s.j.birkinshaw@ncl.ac.uk Z. Rao Halcrow Group Ltd, Burderop Park, Swindon SN4 0QD, UK A rapid assessment method for evaluating the impacts of groundwater abstraction on river flow depletion has been developed and tested. A hybrid approach was taken, in which a neural network model was used to mimic the results from numerical simulations of interactions between groundwater and rivers using the SHETRAN integrated catchment modelling system. The use of a numerical model ensures self-consistent relationships between input and output data which have a physical basis and are smooth and free of noise. The model simulations required large number of input parameters and several types of time series and spatial output data representing river flow depletions and groundwater drawdown. An orthogonal array technique was used to select parameter values from the multi-dimensional parameter space, providing an efficient design for the neural network training as the datasets are reasonably independent. The efficiency of the neural network model was also improved by a data reduction approach involving fitting curves to the outputs from the numerical model without significant loss of information. It was found that the use of these techniques were essential to develop a feasible method of providing rapid access to the results of detailed process-based simulations using neural networks. Key words | artificial neural networks, groundwater, groundwater abstraction impacts, numerical modelling, river–aquifer interactions INTRODUCTION In order to assess the impact of groundwater abstractions on river flows the normal approach is to use either analytical models or numerical models. Analytical solutions (Theis 1941; Hantush 1959; Hunt 1999) are limited in their accuracy or applicability due to the models’ assumptions of homogeneous, isotropic aquifer systems of infinite extent. Many analytical models can represent gaining or losing rivers which are in hydraulic connection with aquifers, but it is difficult to obtain solutions for such nonlinear problems as disconnected rivers, transient (seasonal) recharge inputs or interactions with floodplain wetlands. The usual approach to increase the realism in models is to use a numerical modelling approach. Most numerical models of river – aquifer interaction (see Parkin et al. (2007) for recent examples) involve solution of equations for surface water routing and groundwater flow, with coupling between the two models usually based on a simple Darcy calculation (Winter 1995). However, numerical modelling is both time- consuming and expensive. Hence, an “intermediate” tech- nique is desirable. The approach developed in this study is to use artificial neural networks (ANNs) to mimic numeri- cal model simulations of generic river – aquifer systems, providing a hybrid system which retains the complexity of numerical models and the speed of analytical models. Artificial neural networks are now widely used in hydrology. They have been used in rainfall –runoff modeling (Tokar & Johnson 1999; Anctil & Lauzon 2004; Kumar et al. 2005) and groundwater hydrology (Balkhair 2002; Shigidi & Garcia 2003; Daliakopoulos et al. 2005). The use of hybrid models combining numerical models and ANNs doi: 10.2166/hydro.2008.014 127 Q IWA Publishing 2008 Journal of Hydroinformatics | 10.2 | 2008 Downloaded from https://iwaponline.com/jh/article-pdf/10/2/127/386260/127.pdf by guest on 06 June 2020