Keywords: data supplement, tide, artificial neural network.

The prediction of the tidal variation is an important problem in the planning of constructions or human actives in maritime areas. A complete observation tide database could be helpful to plan the maritime structures. However, the discontinuous observation record may be resulted from the reason of malfunction of facilities, natural disasters or inappropriate operation and so on. This discontinuity could be a few hours or a few months even up to one year. Thus, establishing a simple and executable supplement mode is desired.

In general, the prediction of tidal levels may be obtained satisfactorily, if the numbers of tidal components are sufficient. However, the inclusion of additional unnecessary constituents does not significantly improve the accuracy of prediction. Therefore, appropriate tidal components must be determined priori. As reported in the literature, the determination of the major tidal component using the spectral methods requires at least a year of tidal records for shallow water with significant meteorological noise. In this study, alternatively, we propose to use the corresponding weighting relations in an artificial neural network to determine the tidal components. For a tidal component with significant weighting, its effect is more important for the procedure of prediction. To represent the relations of input layer and weighting function, the back-propagating neural network without a hidden layer is used. In the network structure, 69 tidal components and their corresponding and terms are used, while the tidal levels Y (t) are used in the output layer. Based on one-month and two-month tidal records, the relations between weighting function and the major tidal constituents can be determined based on the neural network model.

Using the determined major constituents, a three-layer of back-propagation neural network is then established for the time series of a tidal-level variation. The interconnection weights among the neurons can be evaluated from the learning processes based on a very short-term observed tidal data. The length of the learning data depends on the length of the missing data to be suppliedt. But less than a 30-day length of learning data is usually sufficient to provide a good performance. In the paper, we back-estimate the missing data in a station during the period from 1994 to 1997. Good agreement between the supplements and the observations are found in the comparison. The results indicate that the artificial neural network is capable of learning the level variations to hindcast or forecast the tidal-level variation using only very short-term observation.

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 £95 +P&P)