12 
Development results 
Tab. 3: Regression table of Phase 2, T + 3 h. 
Predictand Surge, output time T 05 UTC, prediction 
time T 
+ 3 h 
MV 
SD 
R_Pd 
R_Res 
Name 
dRVI 
Co 
Wgt 
Ctr 
10.4 
34.4 
0.942 
0.942 
Last2DModel 
88.6 
0.90 
65 
79 
-3 .3 
12.1 
0.213 
0.630 
Pers2D_korr-2 
39.3 
0.40 
10 
3 
-3 .3 
12.0 
0.188 
0.244 
Pers2D_korr-4 
9.2 
0.29 
7 
2 
5.5 
15.8 
0.583 
0.130 
AbrLastStau 
1.9 
0.14 
5 
4 
-1.0 
8.3 
-0.647 
-0.097 
DMO_DD_SE 
1.5 
-0.26 
-5 
4 
6.6 
4.5 
-0.253 
-0.113 
4Lev_Lift_Ix 
0.9 
-0.21 
-2 
1 
9.8 
33.5 
0.382 
-0.082 
Last2DModel-4 
0.2 
-0.04 
-3 
-1 
5.9 
30.9 
0.767 
0.066 
BrkLastStau-1 
0.6 
0.05 
3 
3 
Const. 
= 0 
4 #Case rm= 
1641 337 RV(HC) = 
94 
SD%(8) 
= 6 
MV (Pd) 
6 
9 #pC eC = 
1611 1641 E(RVI) = 
94 
RMSE 
= 8 
. 7 
SD (Pd) 
= 36 
6 #pPr/Rj = 
253 16 krit_R = 
0.078 
E(RMSI 
= 8 
. 93 
From phase 2 (T+3 hours), DMO is the dominant predictor. It is followed by the predictor Pers2D_korr 
at the time of the last corresponding surge (HW with HW and NW with NW). This predictor includes the 
difference between the persistence of the predictand of the last (-1), next-to-last (-2) etc. surge event 
and the last available model prediction for this particular time. The model's initialisation errors at the 
time of the last corresponding surge are preferably used although they may be 6 hours further back 
than the most recent known initialisation error of the model. This points to its strong dependence on the 
tide (HW or LW). Pers2D_korr accounts for about half of the total reduction of the error variance of 
unclassified MOS as compared to DMO, i.e. for 24% of 46% averaged for all prediction periods. This 
indicates a surprisingly high persistence (temporal autocorrelation from output time to output time) of 
the model's initialisation error. Over selected prediction periods for the 4 observed surge events 
receding the selected output time of 05 UTC, the correlations of the Pers2D_korr predictors are 
distributed as follows: 
Tab. 4: Correlation of the predictors Pers2D_korr-x to the residuum of the 1-predictor equation using the Last2DModel as 
predictor: 
Fp 
-1 
-2 
-3 
-4 
03 
. 16 
. 63 
. 15 
. 58 
09 
.59 
. 16 
. 55 
. 18 
15 
. 16 
. 53 
. 18 
.41 
21 
.50 
. 17 
.39 
. 15 
27 
. 17 
.37 
. 14 
.36 
33 
.36 
. 15 
.36 
. 12 
It is apparent from the dRVI column that by far the largest proportion of the performance of the MOS 
equation is attributable to consideration of this error: decreasing from 46% RV (according to the RV 
addition rules, 39.3% plus 9.2% RV is about 46%) in the equation for T+3 hours to 15% forT+33 hours. 
All other predictors are of lesser importance.