Page 133 - Advances in Renewable Energies and Power Technologies

106 CHAPTER 3 Forecasting of Intermittent Solar Energy Resource

200
persistence measured
180 persistence satellite
cloud motion
cloud motion smoothed
160
satellite
RMSE (Wsq.m) 140
NDFD
120
100
80
60
40
1hour 2hour 3hour 4hour 5hour 6hour 2day 3day 4day 5day 6day 7day

Time Horizon
FIGURE 3.17
RMSE of different solar forecasting techniques obtained over a year at seven SURFRAD
ground measurement sites (from R. Perez, S. Kivalov, J. Schlemmer, K. Hemker Jr., D.
Renne ´, T.E. Hoff, Validation of short and medium term operational solar radiation
forecasts in the US. Solar Energy 84 (2010) 2161e2172). The red (dark gray in print
versions) line shows the satellite now-cast for reference, that is, the satellite “forecast” for
the time when the satellite image was taken. Cloud motion forecasts derived from satellite
[yellow (dark gray in print versions) and white lines] perform better than NWP (NDFD) up
to 5 h ahead. NWP has similar accuracies for forecast horizons ranging from 1 h to 3 days
ahead.
R. Perez, S. Kivalov, J. Schlemmer, K. Hemker Jr., D. Renne ´, T.E. Hoff, Validation of short and medium term
operational solar radiation forecasts in the US, Solar Energy 84 (2010) 2161e2172.

Perez et al. [83] compared several methods of forecasting applied to seven sites;
Fig. 3.17 from [59] and drawn from results obtained by [83] shows the RMSE ob-
tained for each method.

7. THE FUTURE OF THE RENEWABLE ENERGY FORECASTING
Machine learning methods give the best results for the now-casting while methods
based on NWP are the best fit for longer horizons. However, it is reasonable to ima-
gine that in the future, some improvements would make the predictions gain in ef-
ficiency. For the very short term prediction (infra-hour), it is difficult to think that a
major breakthrough could happen: the sky imagers and the ad hoc prediction
methods already give very good results today.

7.1 NOW-CASTING

The literature shows that Support Vector Machine (SVM), ANN, k-NN, regression
tree, boosting, bagging, or random forests systematically give better results than