APPLICATION OF COST FUNCTIONS FOR LARGE SCALE INTEGRATION OF WIND POWER USING A MULTI-SCHEME ENSEMBLE PREDICTION TECHNIQUE - Energy Research Developments: Tidal Energy, Energy Efficiency and Solar Energy - page 349

power of course also has an impact on the error, not only the weather forecast. The dif-
ference of different methodologies for different forecast problems can be quite large. We
will therefore demonstrate this difference in the next section. The following list shows the
methodologies that have been used in this demonstration. The power prediction methods
can be distinguished as follows:
• Method 1: Direction and time independent simple sorting algorithm.
• Method 2: Time dependent and direction independent least square algorithm.
• Method 3: Direction dependent least square algorithm.
• Method 4: Direction dependent least square algorithm using combined forecasts.
• Method 5: Same as method 4, but including stability dependent corrections.
• Method 6: 300 member ensemble forecast of method 5 - all farms are handled indi-
vidually and 6 parameters for each one of the 75 members are used to compute the
power.
Figure 3. Scatter Plot of the mean absolute error (MAE)with a constant background error of
4% added to the native forecast of the EPS mean. The black crosses are EPS mean forecasts
+/- 4% error, while the gray crosses display the errors that lie outside this band.
Note, when considering the results from investigating these different power prediction
methods, all values are given as the improvement in % to the mean absolute error (MAE)
for the day-ahead forecasts based on only one daily (00UTC) forecast performed with the