Numerical modeling in open channel flow using techniques from evolutionary computing.

Abstract

One of the problems in modern society is the contamination of rivers and channels by dangerous chemicals. In order to deal with such problems it is important to predict the transport and diffusion of the contaminants. The way the contaminant plume spreads is not constant along the river, but changes with varying flow conditions. The spread of the plume is mathematically expressed by the diffusion/dispersion coefficients, which consequently should be regarded as spatially varying coefficients. However, with current methods it is not possible to determine these coefficients accurately. They are often taken from a text book and assumed to be constant over the whole river length, although they are varying. This problem can be solved by using techniques from artificial intelligence, which may produce solutions that are superior to those from traditional methods. In particular, genetic algorithms and artificial neural networks are applied to spatially distributed coefficients in one-, two-, and three-dimensional flow models. Using these techniques a large number of spatially varying parameters is found with a very high accuracy compared to measured values---an accomplishment that has not been reached by any other method. This research combines advanced numerical river modeling techniques with the methods of evolutionary computing. The objectives of this research are to (1) create two- and three-dimensional numerical models that adequately describe the flow in open channels, (2) create a hydrodynamic genetic algorithm modeling system, consisting of a two- or three-dimensional numerical model and a genetic algorithm, that can identify spatially varying coefficients in flow and transport modeling, and (3) to investigate the applicability of this modeling system to natural rivers and channels with irregular boundaries. It is found that the application of the modeling system can lead to improved estimations of spatially varying parameters as compared to traditional methods. Genetic algorithms are a valuable method for the calibration of numerical models if the runtimes of the models are short and sufficient field data is available. The robustness of the modeling system makes it especially attractive in applications where other methods do not converge.