Ensemble Based Storm Surge Forecasting Models


Storm surge is a coastal flood or tsunami-like phenomenon of rising water commonly associated with low pressure weather systems such as tropical storms and hurricanes. Accurate prediction of storm surge is a difficult problem. Most forecast systems produce multiple possible forecasts depending on variability in weather conditions possible temperature levels, winds etc. Ensemble modeling techniques have been developed with the stated purpose of obtaining the best forecast (in some specific sense) from the individual forecasts.

With the data from New York Harbor Observing and Prediction System (NYHOPS), we observe those facts:

  • Even the simplest possible way of creating an ensemble produces results superior to any single forecast.
  • For different locations, it may need different model for better forecast
  • We proposed new method in ensemble modling with high accuracy

In this project, we use bayesian learning approach to indetify best ensemble model. Also, we develop some new ensemble forecast by combining the existing forecast models. In order to assign appropriate weights in this combination, we use correlation between each of the forecasts and the observation and standrad deviation of their error to obtain suitable weights. We use these two criterias (correlation and standard deviation) to rank existing forecast models as well.  In addition, we investigate the performance of different ensemble techniques such as Simple Average, Root Mean Square, Bayesian Model Averaging, and Modified Bayesian Model Averaging.   

We select Simple Average (SA) as a benchmark. Then, we use bayesian learning approach to select the best individual forecast. We compare its performance to SA using t-test. We show that a single forecast model is never better than a simple average ensemble model. However, SA is the simplest possible ensemble model. This method is functional when all the forecasts are fine. But, for hurricane events using a simple average may not produce a good ensemble model (Shown in the following two plots).

Also, we investigate whether the performance of ensemble methods is dependent on location and event strength. To achieve this goal, we profrom ANOVA analysis. We find that location has an important effect of the performance of ensemble methods. But, the interaction between event and method is not significant. This analysis helps us to identify best ensemble model for each location. 

  • Amin Salighehdar, Ph.D. Candidate in Financial Engineering (asalighe@stevens.edu)

  • Ziwen Ye, Ph.D. student in Financial Engineering (zye2@stevens.edu)

  • Mingzhe Liu, Ph.D. student in Financial Enginering (mliu13@stevens.edu)

  • Iount Florescu, Research Associate Professor in Financial Engineering Department & Director of Hanlon Financial Systems Lab (ifloresc@stevens.edu)

  • Alan.Blumberg, George Meade Bond Professor & Director of Davidson Laboratory (ablumber@stevens.edu)


1) A.Salighehdar, Z.Ye, M.Liu, I.Florescu,A.F.Blumberg, "Statistical Comparison of Ensemble Based Storm Surge Forecasting Models",  19th International Conference on Coastal and Ocean Engineering (ICCOE 2017). 


2) A.Salighehdar, Z.Ye, M.Liu, I.Florescu, A.F.Blumberg," Ensemble based storm surge forecasting models", Journal of Weather and Forecasting (under review)