Author(s): G. Wallis Steve; Piotrowski Adam; M. Rowinski Pawel; J. Napiorkowski Jaroslaw
Linked Author(s): Pawel M. Rowinski, Stephen Wallis
Keywords: Neural network; Dispersion coefficient; Stream
Abstract: In this paper application of Multi-Layer Perceptron (MLP) Neural Networks to the prediction of dispersion coefficients in a small stream (the Murray Burn in Edinburgh, UK) is presented. Data from eighteen tracer experiments performed in that river are used to test the MLP networks, which were trained and validated on a database of information from other published work. Results from the MLP networks are compared with results from other techniques, such as: the method of moments applied to complete concentration distributions and to distributions with tails truncated at 10% of the peak value; an example empirical formulae and Fischer’s routing procedure. Two diffe rent MLP networks are presented, one trained on all the data in the database and another one that used only data from “smaller rivers”. The performance of the methods was assesse d by comparing the results from each one with those from the routing method. The network trained on “smaller rivers” proved to be the most reliable. Results from the network trained on all the data in the database predicted smaller dispersion coefficients, and they were also smaller than those from Fischer's routing procedure. On the other hand, the method of moments showed the poorest correlation with the routing method. This is caused by a large degree of scatter in the method of moments results, which emanates from it being heavily influenced by the tails of the concentration-time profiles measured in the experimental programme.