Wind power forecasting - Wind Power Integration: Connection and System Operational Aspects - page 225

times from 1 out to 48 hours. As can be seen, and as was mentioned earlier, the
persistence forecast deteriorates rapidly as the look-ahead time is increased.
The MAE measure is easy to interpret directly. It is clear from Figure 6.3 that the
averaged error over the test period for the 6-hour look-ahead time is 16.8 per cent
of installed capacity for the single wind farm and 11.2 per cent for the 15 geo-
graphically dispersed wind farms. The RMSE and SDE measures cannot be inter-
preted so directly as they involve squared errors.
It is also clear that the forecast error measures are reduced for all look-ahead
times for the 15 geographically dispersed wind farms compared to the single wind
farm. This improvement in the error measures expressed in per unit of the single
wind farm value is shown in Figure 6.4. It is very significant at shorter look-ahead
times and drops off as look-ahead time increases. The improvement is due to the
smoothing effects caused by the geographic dispersion of the wind farms.

6.4.2 Reference models
The persistence model is often used as a reference model in wind power forecasting
against which other wind power forecasting models can be evaluated.
A refinement of the persistence model is a moving average (MA) predictor
model. An example of this is where the last measured power value is replaced by
the average of a number of the most recent measured values.

n X

n 1

1

P MA ; n ð t þ k j t Þ¼

P ð t i Þ

i ¼ 0

Although the performance of the moving average forecast model at short
look-ahead times is very poor, it is better than the basic persistence model at longer

0.6

0.5 MAE

RMSE
SDE

0.4

0.3

0.2

0.1

0

0 6 12 18 24 30 36 42 48

Look-ahead time (h)

Figure 6.4 Improvement in basic persistence error measures of 15 geographically

dispersed wind farms over a single wind farm