1.3
Using ensembles of artificial neural networks for storm surge predictions in the North Sea
Daniel Bruce Prouty, Texas A&M University, Corpus Christi, TX; and P. E. Tissot and A. A. Anwar
The use of ensemble forecasting methods utilizing artificial neural network models is evaluated for the prediction of storm surge elevations and arrival times in the North Sea. The models predict storm surges based solely on past measured water level residuals at two tidal stations, the target station, located at Sheerness, England near the entrance of the River Thames, and a secondary station located near the storm at the time of the forecast. The secondary station location is selected to take advantage of the specificity of surge propagation in the North Sea and is located where the peak of the storm surge has just passed. The secondary station location is varied based on the forecast interval. The ANN ensemble models are trained to relate future surge elevations at the Sheerness primary station to the measured surge at a secondary station.
The hypothesis of the work is that the use of ensembles methods can significantly reduce variance and minimize error. The ensemble forecasting method averages results from multiple ANN models trained based on different model initializations. Several different models are tested, varying the number of members used for each ensemble as well as the number of neurons used for individual members. The ensemble performances are analyzed on an annual basis and on a 72-hour window centered on individual storm events, which focuses the evaluation on a time when it is most critical.
The use of ensemble forecasting with ANNs is found to significantly reduce variance when analyzed over a 72-hour storm window, but not model accuracy. The average absolute error for an ensemble ANN using 5 repetitions has 50% of the variance of a single ANN model. An ensemble model using 50 repetitions has 5% of the variance of a single ANN model. A significant result of this research is the ANN's ability to accurately predict maximum water elevations at the project target station. A single ANN model has a 4-hour forecast error of 0.017 m, while a simple [1,1] ensemble model using 20 repetitions performs better with an average 4-hour forecast error of 0.008 m. When over-training is included to reduce the model bias, the error is further reduced to 0.004 m. ANN ensemble model performances for predicting tide elevation during the time of maximum surge were however less impressive. Best results were obtained for ensembles of [30,1] models with an average 4-hour forecast error of 0.68 m.
Recorded presentationSession 1, Application of Artificial Intelligence to Storm Surge Forecasting
Monday, 21 January 2008, 4:00 PM-6:00 PM, 205
Previous paper Next paper