Computational & Technology Resources
an online resource for computational,
engineering & technology publications
Civil-Comp Proceedings
ISSN 1759-3433
CCP: 78
Edited by: B.H.V. Topping
Paper 57

A Neural Network Approach for Seepage Control in Earth Dams

J. Veiga Carvalho, J.L.C. Gutiérrez and C. Romanel

Department of Civil Engineering, Catholic University of Rio de Janeiro, Brazil

Full Bibliographic Reference for this paper
, "A Neural Network Approach for Seepage Control in Earth Dams", in B.H.V. Topping, (Editor), "Proceedings of the Seventh International Conference on the Application of Artificial Intelligence to Civil and Structural Engineering", Civil-Comp Press, Stirlingshire, UK, Paper 57, 2003. doi:10.4203/ccp.78.57
Keywords: water discharge, dam monitoring, time series, drainage, Funil dam, Corumba-I dam.

Dams are monitored in order to verify whether their performance is consistent with the original design expectations. Performance monitoring of dams is generally accomplished by reviewing and analysing data collected from instruments which measure critical indicators of engineering behaviour. Instrumented monitoring includes measurements of displacement, strain, stress, pressure, loads on structural members, seepage and drainage along with environmental factors that affect the dam behaviour such as temperature, reservoir level and precipitation.

In this work, artificial neural networks (ANN) and the Box Jenkins approach [1] are used to model time series consisted of a sequence of water discharges, measured along several years of continuous observation, through the foundation of two large Brazilian dams (Funil and Corumba-I dams). The sequence of water discharges through the soil foundation were modelled using multi-layer feedforward networks. The algorithm for ANN training was either the LM (Levenberg-Marquardt) method, for the Funil dam time series, or the backpropagation algorithm, with the descendent gradient method, for the data sets collected at the Corumba-I dam. Several neural network configurations, with different numbers of input parameters and hidden neurons, were trained and tested to assess their influence on the model behaviour.

The time series for Corumba-I dam consisted of 130 patterns, measured at time intervals of 15 days, beginning in March 1997 and extending to December 2002. Two data sets were considered in this analysis: the first one composed of 104 samples, used during the training phase, and the remaining 26 patterns, which form the second data set, reserved for the validation process. Figures 57.1 and 57.2 compare the actual measured values with respective neural network forecasts for the case of Corumba-I dam.

The computed results are quite helpful to control seepage under the dam foundation, since a study with conventional methods would require a 3D finite element model and a comprehensive field investigation in order to determine the local soil properties which, in this particular area, presents a system of fractures that rends the analysis a quite difficult task. The ANN performances were considered better than those obtained by the AR [1] model. However, the Box-Jenkins approach was a helpful tool for preliminary analyses, which permitted an a priori estimate of the required number of sequence elements for the ANN models.

Figure 57.1: Comparison between real and forecasted values during the training phase of the neural network for Corumba-I dam.

Figure 57.2: Comparison between real and forecasted values during the validation phase of the neural network for Corumba-I dam.

Box, G.E.; Jenkins, G.M., "Time Series Analysis", Holden-Day, San Francisco, 1970.

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 £82 +P&P)