85th AMS Annual Meeting

Monday, 10 January 2005: 2:00 PM
Using Artificial Neural Networks for Predicting Storm Surge Propagation in North Sea and the Thames Estuary.
Daniel B. Prouty, University of Southampton, Southampton, United Kingdom; and P. E. Tissot and A. A. Anwar

The United Kingdom's economy relies heavily on waterborne commerce and its ports system is one of the largest in the European Union. Several of the ports are located along the North Sea coast and within the Thames River Estuary. Water level forecasts are important to the shipping industry using these facilities and waterways especially considering the continuing worldwide increase in vessel draft. Forecasting the occurrence and extent of high water level events is also important for coastal users, industries and for the city of London, as 150km of London lies below mean high spring tide level. Due to recurrent flooding, a system of barriers was constructed and brought on line in 1983 to protect the city and navigational concerns along the lower River Thames. The system is enabled when extreme high or Low water levels are predicted.

Presently there are several methods for predicting water levels. The most common method uses tide tables, which are computed by harmonic analysis of previous water level records and is therefore strictly based on astronomical forcing. The method works well for large portions of the year but meteorological effects can significantly influence water levels and introduce substantial errors in the predictions. Large variations from the tidally forecasted water levels as high as 2 meters are observed during frontal passages and large storm events. Other methods developed for predicting accurate water levels include the persistence model and finite element models. Currently in the United Kingdom, extreme water levels are predicted by the Storm Tide Forecasting Service (STFS). Recently, development of Artificial Neural Network (ANN) models has enabled accurate water level predictions by including meteorological effects. The main advantages and key characteristics of neural networks for water level forecasts are their non-linear modeling capacity, their generic modeling capacity, their robustness to noisy data, and their ability to deal with high dimensional data.

The study focuses on developing and assessing a model to forecast water levels at the Sheerness station at the entrance of the River Thames based on Artificial Neural Networks (ANN's). The ANN model is optimized for Sheerness based on selection of a secondary water level measurement station location , selection of training year, the number of previous water level measurements, and forecast time interval. The model takes advantage of the relatively consistent southward progression of North Sea storms to predict the future surge height based on past water level measurements at both the Sheerness station and the secondary northern station.

The ANN model performance is compared with harmonic and persistence forecasts. The comparisons are quantified based on the root mean square error (R.M.S.E.) and the central frequency of 15 cm or the percentage of forecasts which are within 15 cm of the actual measurement. Models predicting future water levels for the Sheerness station are optimized for 6, 12, and 24-hour forecasts. The composition of the input deck, the number of hidden neurons and the location of the secondary northern station are varied during the optimization process. For example, a simple ANN model computing 6-hour water level predictions using 48 hours of previous water level measurements at the Sheerness and Immingham stations was trained over 1991 water level records. The performance of this ANN model when applied to the 1992 water level records yields a R.M.S.E. of 11.7 cm as compared with 28.5 cm for the persistence model and 25.0 cm for harmonic predictions. For the same case, the central frequency for the ANN model is 86% as compared to 43% for the persistence model and 51% for the harmonic predictions. The overall performance of the ANN models and their operational potential is discussed. The new ANN models show promises for emergency management, the shipping industry and overall management of coastal and riverine activities and resources.

Supplementary URL: