Development results 
13 
Tab. 5: Regression table of Phase 2, T + 33 h. 
Predictand Surge, output time T 05 UTC, prediction time T + 33 h: 
MV 
SD 
R_Pd 
R_Res 
Name 
dRVI 
Co 
Wgt 1 
Ctr 
10.9 
37.9 
0.915 
0.915 
Last2DModel 
83.6 
0.78 
64 
74 
-3.4 
12.2 
0.148 
0.365 
Pers2D_korr-3 
12.8 
0.24 
6 
1 
16.9 
10.4 - 
-0.085 
0.159 
FI_OCisotherm 
2.2 
0.19 
4 
0 
25.2 
102.9 
0.750 
0.097 
WStLS_295* *2 
1.7 
0.04 
8 
8 
-4.1 
23.8 
0.145 
0.140 
Rot_1000/FF1000 
1.5 
0.09 
5 
1 
-3.4 
12.3 
0.145 
0.125 
Pers2D_korr-1 
2.0 
0.21 
6 
1 
27.3 
31.9 
0.524 
0.089 
U_300 
0.5 
0.06 
4 
3 
-6.6 
70.1 - 
-0.057 - 
•0.093 
Cos_2*Dag 
0.3 
-0.02 
-3 
0 
Const. 
= -5. 
.9 #Case rm= 
1648 330 RV(HC) = 
87 
SD%(8) 
= 7 
MV(Pd) 
6 . 
.6 #pC eC = 
1621 1648 E(RVI) = 
87 
RMSE 
= 12 
. 8 
SD(Pd) 
= 36. 
.3 #pPr/Rj = 
259 16 krit_R = 
0.078 
E(RMSi; 
)= 13 
. 15 
In the two equations, MOS reduces the error variance of DMO by about 50% (T+3h) and 35% (T+33h), 
see Table 7 below. 
Classification of predictands 
Experience gained in oceanographic operational applications has shown that HW and LW surges react 
differently to otherwise identical surge signals; accordingly, positive and negative surges are governed 
by different "rules" (prediction rules and experience). This led to the novel approach of making a 
classification based on the two criteria, initially 2x2 classes with HW and LW or surge+ and surge-, 
respectively, accounting for about half of all cases in a particular class; later and in operational use even 
4 classes: HW+, HW-, LW+, LW-. 
Averaged overall prediction periods analysed, an even higher reduction of variance has been achieved 
with classified predictands (28%) than by making allowance for the initialisation error of the 2Dv4 
model with an unclassified predictand (26%). The improvement is mainly due to a more differentiated 
consideration of the role played by the inialisation error of DMO. For, consideration of the initialisation 
error is much more successful with negative surges than with positive surges, as can be seen in the 
Table 6 below: 
Table 6: Reduction of the error variance of the Pers2D_korr predictors for the four classes, in % 
LW+ HW+ LW- HW- 
8 19 39 52 
MOS improvements to DMO thus are likely to be most successful with expected negative surges, 
particularly at high tide. Such results should be of interest to modellers looking for ways of improving 
their models. However, in order to preserve the quality of the DMO+MOS overall system, they should 
avoid changing their v4 models but, instead, use these diagnoses in developing a next v5 model 
version. Table 7 summarises the results obtained so far in verifying the development system - with 
regard to the importance of different groups of predictors - for selected prediction periods Fp: