Wind power forecasting - Wind Power Integration: Connection and System Operational Aspects - page 240

Sub-area 1 Sub-area 2 Sub-area 3 Total area

Meteorological
service
NWP

WPPT

Data check Data check

Data check

Model Model Model Wind farm prediction

Sub-area prediction

Upscaling Upscaling Upscaling

Sum Total area prediction

GUI

User 2

User 1 User 3

Figure 6.20 Schematic of WPPT wind power forecasting model

time series approaches for look-ahead times up to 10 hours, (ii) longer-term fore-
casts based on F-NNs with inputs from SCADA and NWP for look-ahead times out
to 72 hours and (iii) combined forecasts produced from an intelligent weighting of
short- and long-term forecasts to optimise the performance over all look-ahead
times. The results shown in Figures 6.22 and 6.23 are taken from Pinson et al.
(2004) and are for a single offshore wind farm of 5 MW capacity and are based on a
13-month set of data which was divided into test and training subsets. Figure 6.22
shows the comparison with persistence using two different normalised error mea-
sures - the MAE and the RMSE. The typical pattern is again to be seen, with
AWPPS outperforming persistence, as would be expected. The results for a single
offshore wind farm are clearly not as good as for a whole region. In Figure 6.23 the
improvement obtained over persistence is plotted against look-ahead time.
The largest source of error for a wind power forecasting system is in the NWP.
It is important to indicate to users the uncertainty that is attached to a particular
forecast. Figure 6.24, again taken from Pinson et al. (2004), shows a sample plot of
forecast power with upper and lower 85 per cent confidence intervals and the
measured power plotted against look-ahead time. Tools for the online estimation of