Author(s): Jaroslaw J. Napiorkowski; Adam P. Piotrowski; Pawel M. Rowinski; Steve G. Wallis
Linked Author(s): Pawel M. Rowinski
Keywords: No Keywords
Abstract: The problem of estimating longitudinal dispersion coefficients in rivers, although studied for decades, is still a difficult task. A number of empirical equations have been proposed, many of them in a multiple power law regression form. Also, during the last ten years a number of datadriven techniques have been suggested to improve the results, including a few types of neural networks. However, Product-Unit neural networks (PUNNs), which should be well suited for dispersion prediction, have never been used for this task. Hence, in this paper PUNNs are applied to estimate longitudinal dispersion coefficients in rivers. As identifying the global optimum of PUNNs is much more difficult than for classical Multi-Layer Perceptron neural networks (MLPs), two different global optimization training algorithms are compared. In order to avoid the problem of overfitting of neural networks to training data the popular noise injection method is used. Based on 50training runs, average objective function values from the PUNNs are generally not as good as those from the MLPs. However, if the best of the 50 runs are considered, the PUNNs allow for slightly better objective function values than MLPs. In general, noise injection makes a significant improvement to DEGL trained MLPs, but it appears to be much less beneficial for PUNNs.