Session of the wmo-ioc joint Technical Commission for Oceanography and Marine Meteorology (jcomm) agreed that it would be logical to transform the wmo wave Programme into the jcomm wind Wave and Storm Surge Programme

Harris (1962) appears to be among the first who suggested the relevance of statistical methods for storm surge prediction.

Figure 1.5: Locations of grid points (+) and observing stations () for the Great Lakes area. (Venkatesh 1974).

Figure 1.6 Multi-layer perception structures with one hidden layer and a hidden unit

Table 1.2 outlines several experimental setups and Tables 1.3–1.6 list the results of these model experiments. The results indicate the accuracy of the neural network model forecasts. The root mean square error (RSME) between MLP and RBF differed only slightly, but MLP was more stable than RBF. The best results here were obtained when MLP training involved three hidden units and previous input data, to a RMSE of approximate 20 cm, which is about 5% of the tidal range at Jeju and Yeosu. Figure 1.8 presents the results of Model 1 for Busan station using the data predicted by the Mesoscale Model version 5 from the Korea Meteorological Administration (MM5–KMA) on September 12–13, 2003. Table 1.7 presents the performance of Model 2 for Yeosu station data for forecasting times of 3, 6, 12, and 24 h. For 24-h forecasts, root mean square error improves from 13–39 cm to 12–15 cm.

Table 1.2 Summary of experiments used to determine optimal neural network model structure

Table 1.3 Root mean square error of the models from experiment 1 and harmonic forecast at Jeju and Yeosu station on the Korean Peninsula (in cm)

Table 1.4 Root mean square error of the models from experiment 2 and harmonic forecast at Jeju and Yeosu station on the Korean Peninsula (in cm)

Table 1.5 Root mean square error of the models from experiments 1 and 3 and harmonic forecast (in cm)

Table 1.6 Results of experiments 4 and 5 : Comparison between observed sea levels and forecasts at Yeosu station during Typhoon Maemi (in cm)

Table 1.7 Model performance for typhoons at Yeosu station on the Korean Peninsula : Model 2

Figure 1.7 Results of experiment 5 : Observed sea level (solid line) and predicted sea levels from HA (dashed) and ANN (dotted) at Yeosu station during the passage of (a) Typhoon Gladys in 1991, (b) Faye in 1995, (c) Yanni in 1998 and (d) Rusa in 2002

Figure 1.8 Observed sea level, sea levels predicted by harmonic analysis and artificial neural network for Busan station on the Korean Peninsula : Model 1.

These results indicate that it is possible to apply a neural network model for each cluster to the regional storm surge forecasting system and to extend this into all coastal regions. Extension into all coastal regions requires continual updates of typhoon meteorlogical data. In addition, model accuracy can be improved by correcting predicted air pressure and wind stress, which uses model input variables. Future research will include construction of an autonomous modeling system with a real-time training optimization model that includes the present meteorological state as well as past typhoon information.

Figure 1.9 Real-time storm surge prediction using the neural network model at Busan, Yeosu and Wando stations on the Korean Peninsula. (Left panel: Observed (dotted) and predicted sea level (dashed), predicted tides (straight), Right panel: Predicted storm surges)

7. 
   Numerical Methods
   

Numerical methods are now the widely used approach for storm surge prediction. For a review of such methods, see Gonnert et al. (2001). Until about the 1980s, finite-difference models with rectangular (or square) grids were traditionally used to get better resolution near the coastlines; telescoped grids, such as shown in Figure 1.9 for the Bay of Fundy were used (Greenberg 1979).

Figure 1.9: Telescoped rectangular grids for modeling tides in the Bay of Fundy (Greenberg 1979).

Figure 1.10: The East Coast, Gulf of Mexico, and Caribbean Sea computational domain. The scale shown in the legend is in Km (Scheffner et al. 1994)

Change of sea level due to the static influence of atmospheric pressure is usually referred to as the inverse barometer effect. According to this, a pressure change of 1 mbar causes a sea level change of approximately 1 cm. When pressure systems are moving with a definite speed, the sea level at a point does not correspond to the static value and under certain conditions they can be significantly increased due to resonance.

Water level oscillations occur predominantly at the fundamental natural period of the basin and they depend less on the temporal scale of the weather system. When forcing exceeds the natural period of the system, water level oscillations with various natural frequencies with uniform distribution of energy in the perturbation spectrum are observed after a short duration external impulse is applied.

Tangential stress of the wind per unit area is usually expressed through an average wind speed W, air density and air turbulence coefficient k_a:

 = 0.4 (the von Karman constant)

z₀ – aerodynamic roughness of underlying surface

V – traverse speed of underlying surface

_a – air density

W – surface wind speed.
In general, C_D is not a constant coefficient, but is a varying function depending on the degree of surface roughness, atmosphere stratification and other wind parameters. There are many good reviews that discuss determining the drag coefficient (e.g. Parker, 1975; Smith and Banke, 1975) and more recent papers (e.g. Powell et al., 2003) address constraints on drag coefficients at the very high wind speeds found in tropical cyclones. Nevertheless C_D is often taken as constant in many models of storm surges. Such an approach may be valid when modeling short-duration surges covering relatively small areas of the coast (i.e. where the tangential stress of the wind field on the sea surface has little variability in space and time). However, for long duration storm surges over extensive stretches of a coastline one should take into consideration the spatial and temporal variability of the drag coefficient.



Figure 2.1: Friction coefficient calculated for different wind speeds over a water surface and measured at different stratifications. 1: Brox (Roll, 1968), 2, 3: Deacon (Deacon, 1962, Deacon and Webb 1962); 4: Deacon’s formulae c^s = (1.1 + 0.04W) x10^-3; 5: c^s = (1.5 + 0.01W) x10^-3
In coastal areas (or estuaries or lakes) wind speeds from an atmospheric model are, on average, lower over land than over water. This physical fact, coupled with resolution of the atmospheric and surge models, can give an underestimate of the wind speed used for surge modelling. Surface (or 10m) wind speeds are usually derived in a post-processing step after running the atmospheric model itself. A boundary-layer wind profile between the surface and the lowest model level is determined. Where the surface type changes within a grid box, some averaging has to be applied.
To overcome this problem without increasing the resolution of the atmospheric model, a technique called downscaling can be used (Verkaik et al., 2003; Verkaik, 2006). The technique uses a high-resolution roughness map and a simple two-layer model of the planetary boundary layer (PBL). The roughness map is determined from a land-use map and a footprint model to account for upstream properties of the surface. The roughness lengths are hence dependent on the wind direction. The PBL model is first used to calculate the wind at the top of the PBL from the surface wind from the atmospheric model with the roughness length which is used by the atmospheric model. In a second step the surface wind is recalculated with the high-resolution roughness map.