Artificial Neural Network Design and Optimization for the Prediction of Observed Tidal Levels

Observed sea levels can vary greatly from the levels predicted by the harmonic constants method. This variation is largely the result of meteorological effects combined with the topography of the adjacent ocean basin and is commonly known as storm surge. Artificial neural networks are being increasingly used to model complex relationships and have recently been used to improve the accuracy of the tidal predictions. This investigation describes the process of designing, and optimizing, an artificial neural network of the multilayer perceptron type, for the prediction of observed sea levels. The effects of network architecture, and internal parameter selection on the network performance, are also described. The results obtained indicate that the learning rate, momentum, noise level and number of epochs does not have a significant effect upon network performance. The number of nodes in the hidden layer did appear to affect the network performance, but the limitations of the number used do not allow firm conclusions to be made. The best network developed in this investigation was found to have a poor performance. The proposed reasons for this are that: the model lacks a suitable level of complexity; the data set has an insufficient volume of cases; and the selection of input variables was insufficient. A number of possible improvements to the technique set out in this study have been identified.Faculty of Scienc