Understanding the characteristics of probability distribution of extreme water levels is important for coastal flood mitigation and engineering design. In this study, frequency analysis has been conducted to investigate probability distributions along the coast of the U.S. by using three-parameter General Extreme Value (GEV) method. The GEV model combines three types of probability distributions (Type I for Gumbel distribution, Type II for Fretchet, or Type III for Weibull) into one expression. Types of distributions can be clarified by one of the three parameters of the GEV model for the corresponding studied stations. In this study, the whole U.S. coast was divided into four study areas: Pacific Coast, Northeast Atlantic Coast, Southeast Atlantic Coast and Gulf of Mexico Coast. Nine National Oceanic and Atmospheric Administration (NOAA) stations with a long history of data (more than 70 years) in the four study areas were chosen in this study. Parameters of the GEV model were estimated by using the annual maximum water level of studied stations based on the Maximum Likelihood Estimation (MLE) method. T-test was applied in this study to tell if the parameter, , was greater than, less than or equal to 0, which was used to tell the type of the GEV model. Results show that different coastal areas have different probability distribution characteristics. The characteristics of probability distribution in Pacific Coast and Northeast Atlantic Coast are similar with extreme value I and III model. The Southeast Atlantic Coast and Gulf of Mexico Coast were found to have similar probability distribution characteristics. The probability distributions were found to be extreme value I and II model, which are different from those of the Pacific Coast and Northeast Atlantic Coast. The performance of the GEV model was also studied in the four coastal areas. GEV model works well in the five studied stations of both the Pacific Coast and the Northeast Atlantic Coast but does not work well in the Southeast Atlantic Coast and the Gulf of Mexico Coast.
Adequate predictions of extreme annual maximum coastal water levels (such as 100-year flood elevation) are also very important for flood hazard mitigation in coastal areas of Florida, USA. In this study, a frequency analysis method has been developed to provide more accurate predictions of 1% annual maximum water levels for the Florida coast waters. Using 82 and 94 years of water level data at Pensacola and Fernandina, performances of traditional frequency analysis methods, including advanced method of Generalized Extreme Value distribution method, have been evaluated. Comparison with observations of annual maximum water levels with 83 and 95 return years indicate that traditional methods are unable to provide satisfactory predictions of 1% annual maximum water levels to account for hurricane-induced extreme water levels. Based on the characteristics of annual maximum water level distribution Pensacola and Fernandina stations, a new probability distribution method has been developed in this study. Comparison with observations indicates that the method presented in this study significantly improves the accuracy of predictions of 1% annual maximum water levels. For Fernandina station, predictions of extreme water level match well with the general trend of observations. With a correlation coefficient of 0.98, the error for the maximum observed extreme water level of 3.11 m (NGVD datum) with 95 return years is 0.92 %. For Pensacola station, the prediction error for the maximum observed extreme water level with a return period of 83 years is 5.5 %, with a correlation value of 0.98.
In frequency analysis of 100 year coastal flood (FEMA 2005), annual extreme high water levels are often used. However, in many coastal areas, long history data of water levels are unavailable. In addition, some water level records may be missed due to the damage of measurement instruments during hurricanes. In this study, a method has been developed to employ artificial neural network and harmonic analysis for predicting extreme coastal water levels during hurricanes. The combined water levels were de-composed into tidal signals and storm surge. Tidal signal can be derived by harmonic analysis, while storm surge can be predicted by neural network modeling based on the observed wind speeds and atmospheric pressure. The neural network model employs three-layer feed-forward backgropagation structure with advanced scaled conjugate training algorithm. The method presented in this study has been successfully tested in Panama City Beach and Apalachicola located in Florida coast for Hurricane Dennis and Hurricane Ivan. In both stations, model predicted peak elevations match well with observations in both hurricane events. The decomposed storm surge hydrograph also make it possible for analysis potential extreme water levels if storm surge occurs during spring high tide.