Sedimentation in the Maasmond: Non-linear statistics in Civil Engineering

Sedimentation in the Maasmond: Non-linear statistics in Civil Engineering

Bierens, R.

Schiereck, G.J. (mentor)
Vrijling, J.K. (mentor)
D' Angremond, K. (mentor)

Civil Engineering and Geosciences

Hydraulic Engineering

1997-08-29

Rijkswaterstaat and the Dutch contractors have many years of experience in the removal of silt and sand from the Rotterdam harbour. In all those years a great deal of valuable experience has been gained with the behaviour of silt in the harbour. They never succeeded, however in making an estimation/prediction of the amount of sedimentation which has taken place, based on natural parameters like wind river discharge etc.. In this study an attempt is made to build a Neural Network model in order to perform an adequate simulation. A Neural Network is a powerful nonlinear data analysis tool; the network developed for the sedimentation simulation is called Mud Brain. In order to perform the Neural Network analysis it is necessary to make a re-analysis of the existing measurement data. It is also necessary to make an inventory of the knowledge available about the sediment transport to the Maasmond, starting off with the large scale sediment transports in the Southern North sea. A silt and sand transporting residual-current runs parallel to the Belgian and Dutch shoreline, it is mainly driven by the tide but is strengthened by the ruling wind direction which is from the south-west. Sediment supplied through the Channel and eroded from the Vlaamse Banken travels to the to the Wadden Sea and Oester Gronden. While on transport, sediment sometimes settles under tranquil weather conditions, however, most of the sediment is eroded again and travels further north. Due to the unnatural depth, a great deal of sediment is caught in the Maasmond. The material which enters mostly settles in dredging area F. About 40 % of the material is further transported further along to dredging area E (the buffer pit). In their present state, numerical models are not yet able to predict the amounts of sediment settling in the Maasmond, based on certain natural parameters. The best option for a sediment prediction seems to be a statistical analysis on the relation between the amount of sediment and some natural parameters. For most patterns information about the past as well as information about the present is included (except for the first two patterns which already contain a hitorical time span). In order to provide the statistical analysis with the best possible target outputs, a new sedimentation model is drawn up. This model uses the 1.03-level, the 1.2-level, the amounts of dredged material and some consolidation theory in order to calculate the quantities of sediment, which have settled in a certain dredging area, for every week. This reports presents a test for two statistical methods: 1. Linear regression and 2. Neural Network processing. The Neural Network produces results with a 50 % smaller error margin than the linear regression. The Neural Network results are produced with the so-called 'Leave One Out' testing procedure. This procedure is applied to include all available patterns in the test procedure. After post processing the results of the LOO test fall within an accuracy margin of 23,000 TDS and 34 %. When the standard deviation on the amounts of sediment (= 34,000 TDS, target outputs of the model) are taken into account this result seems adequate. The Levenberg-Marquadt algorithm has proven to produce faster and more accurate results in the sedimentation prediction than conventional back propagation. Recurrent networks and data reduction provide valuable tools in making initial estimations of the attainable accuracy and the important parameters. The most promising potentials of Mud Brain are: • An analysis of the relative importance of the different input parameters on the amount of sedimentation • The application of Mud Brain for a sedimentation prediction in order to smooth the dredging logistics. • The combination of Neural Networks and mathematical physical models in a hybrid model. In this combination the neural network can function as a filter on the mathematical-physical model results.

neural network
sedimentation
river discharge
siltation

http://resolver.tudelft.nl/uuid:6037a3ea-a3af-4d04-97b4-cbb9d482a1eb

Student theses

master thesis

(c) 1997 Bierens, R.