Development results 
11 
4 Development results 
There are two basically different forecast phases. 
Tab. 2: Regression table of Phase 1, T+1 h. 
Predictand Surge, output time T 05 UTC, prediction 
time T 
+ 1 h: 
MV 
SD 
R_Pd 
R_Res 
Name 
dRVI 
Co 
Wgt 
Ctr 
5.9 
30.9 
0.970 
0.970 
BrkLastStau 
94.1 
0.77 
51 
63 
208.8 
587.3 
0.769 
0.480 
WStl_295* *2 
41.8 
0.01 
9 
9 
10.2 
34.2 
0.943 
0.072 
Last2DModel 
5.1 
0.34 
25 
30 
-3.3 
12.1 
0.213 
0.322 
Pers2D_korr-2 
14.0 
0.17 
4 
1 
5.9 
30.8 
0.765 - 
0.140 
BrkLastStau-1 
5.1 
-0.07 
-5 
-5 
111. 6 
89.6 
-0.328 
0.194 
FI_ 1000 
4.3 
0.02 
3 
-1 
-3.3 
12.0 
0.185 
0.142 
Pers2D_korr-4 
3 . 0 
0.12 
3 
1 
Const. 
= -2 
9 #Case rm= 
1639 340 RV(HC) = 
98 
SD%(8) 
= 7 
MV(Pd) 
6 
9 #pC 
eC = 
1609 1639 E(RVI) = 
98 
RMSE 
= 5 
. 6 
SD(Pd) 
= 36 
5 #pPr/Rj = 
254 17 krit_R = 
0.078 
E(RMSI 
= 5 
. 74 
The last HW or LW surge at Borkum - determined not later than one hour before output time - is an 
excellent indicator of the surge level to be expected at Cuxhaven. Being the best-correlated predictor, 
it is selected first by the regression algorithm. The dRVI column contains the values of the expected 
reduction of the error variance due to integration of the particular predictor into the equation. 
In the case of the first predictor, this value is 94.1% of the initial variance of the predictand of 
SD(Pd)2=(36.5 cm)2. The error of the 1 -predictor equation with BrkLastStau is corrected most effectively, 
with the error variance reduced by more than 40%, by wind surge computed on the basis of the square 
of GFS-predicted geostrophic 1000 hPa wind from a 295° direction (WNW). The next predictor is the 
last 2D model and the correction of its last known initialisation error, which has been discussed in more 
detail in the explanations to phase 2 below. The next-to-last surge in Borkum (BrkLastStau-1) is used 
with a negative sign; together with BrkLastStau, it forms a trend equation. A further 4.3% reduction of 
the remaining error variance is achieved by the geopotential in 1000 hPa, and another 3% by correcting 
the model's initialisation error for the next-to-last corresponding surge event. It is apparent from the 
Wgt column, which shows the predictor weighting standardised to 100% (with the sign of the regression 
coefficient Co), that BrkLastStau accounts for well over 50% of the total reduction of the error variance 
of this equation, followed by DMO, at 25%, and GFS wind surge at just under 10%. The other predictors 
account for the remaining 15% of weighting. The MOS equations for T+2 hours are of comparable quality. 
The weight percentages add up to 100%, the values in the Ctr column (contribution) to RV=98%, and 
dRVI according to the above formula: 
RVI = (1 - Y\U ~ dRVI, /100%]) ■ 100% (2) 
In equation (2) dRVI. designates the reduction of the error variance due to integration of the new i th 
predictor.