Computational & Technology Resources
an online resource for computational,
engineering & technology publications
Civil-Comp Proceedings
ISSN 1759-3433
CCP: 92
PROCEEDINGS OF THE FIRST INTERNATIONAL CONFERENCE ON SOFT COMPUTING TECHNOLOGY IN CIVIL, STRUCTURAL AND ENVIRONMENTAL ENGINEERING
Edited by: B.H.V. Topping and Y. Tsompanakis
Paper 11

Predicting the Change in Hydraulic Head of a Karstic Aquifer using Neural Networks

I. Trichakis1, I.K. Nikolos2 and G.P. Karatzas1

1Department of Environmental Engineering, 2Department of Production Engineering and Management,
Technical University of Crete, Greece

Full Bibliographic Reference for this paper
I. Trichakis, I.K. Nikolos, G.P. Karatzas, "Predicting the Change in Hydraulic Head of a Karstic Aquifer using Neural Networks", in B.H.V. Topping, Y. Tsompanakis, (Editors), "Proceedings of the First International Conference on Soft Computing Technology in Civil, Structural and Environmental Engineering", Civil-Comp Press, Stirlingshire, UK, Paper 11, 2009. doi:10.4203/ccp.92.11
Keywords: artificial neural networks, differential evolution, hydraulic head prediction, karstic aquifers.

Summary
In the present work, artificial neural networks (ANNs) are utilized to predict the response of a karstic aquifer. The objective of this work is to support the construction of ANNs with the knowledge of the physical system and provide physical meaning to the parameters used as input and output parameters of the ANNs, in order to overcome their "black box approach". Two different multi layer perceptron ANNs were utilized; the first one is used to predict the hydraulic head at two observation wells, while the second one was modified to predict the change to the hydraulic head between two successive days.

The area of study is the pumping site of Mavrosouvala, in the North West part of the prefecture of Attica in Greece. The geology of the aquifer consists mainly of limestone, while its recharge comes from the mountains in the south. The karstic system of the region is not accurately described by geophysical analysis, so it is impossible to build a groundwater model using the dual porosity concept or pipe flow theory. The water table is rather low, about 100 m below the ground surface and 15-20 m above mean sea level (MSL). In the area of study three measuring stations provide rainfall measurements, sixteen wells are used for pumping and two observation wells provide hydraulic head measurements.

The ANN adopted is a classic fully connected multilayer perceptron, trained in a supervised manner with the error back-propagation algorithm. Two hidden layers were used and the activation function is the commonly used logistic function. The input parameters used for the ANN model are: the current day number, the current day temperature, the rainfall measurements from three measuring stations some days before the current one, the pumping duration for the sixteen pumping wells, and the hydraulic head at the two observation wells the previous to the current day.

The correlation of rainfall and hydraulic head change was initially used to determine the time lag of the rainfall input data, representing the time needed by the rainfall to reach the water table. The correct determination of the time lag between the current day and the day used for input to the ANN of the measured rainfall levels proved to be critical in the correct modelling of the aquifer's aquatic equilibrium. The simulation results showed that the first network failed to provide a good approximation of the physical phenomenon, while the second one showed a significantly improved behavior. The reason is that the hydraulic head change than the hydraulic head itself is related to the aquatic equilibrium and the ANN's input parameters. As a result, an improved association between input and output parameters can be achieved and the network provides better physical meaning. It is therefore suggested that, when groundwater simulation takes place using ANNs, the hydraulic head change should be used instead of the hydraulic head.

A differential evolution algorithm was also used to optimally define the time lag in the rainfall measurements as well as the ANN's architecture and training parameters. A significant improvement in the behavior of the first network was achieved, while for the second one the improvement was marginal, as the initial results were already very good. As the optimization procedure was time consuming, it is questionable if such a procedure can be easily used for practical applications. However, the improvement obtained for the first network suggests that the empirical determination of the ANN parameters and structure is not always sufficient.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £78 +P&P)