Page 134 - Advances in Renewable Energies and Power Technologies

7. The Future of the Renewable Energy Forecasting 107

classical regression methods [92]. All the methods based on the use of regression
trees or similar methods (boosting, bagging, or random forest) are barely used
even though they give excellent results. It is not easy at this stage to draw a conclu-
sion, but in the next 5 years, it is probable that these methods might become a refer-
ence in terms of irradiation prediction. An alternative to all the previous methods is
certainly k-NN, but this should be proved. Actually, it is very difficult to propose a
ranking for machine learning methods, although SVR, regression trees, boosting,
bagging, and random forest seem the most efficient ones. To overcome this problem,
some authors do not hesitate to combine single predictors. There are a lot of solu-
tions that combine predictors. Often ANN is used to construct the predictors [56].
Systematically, the ensemble of predictors gives better results than a single predictor
but the best methodology of hybridization is not really defined. Further works are
necessary to propose a more robust method, or maybe, to prove that all the methods
are equivalent. For the moment, it is not demonstrated! Several methods exist to es-
timate the solar radiation. Some of them are often used such as ANN, ARIMA, and
naı ¨ve methods; others are begin used more frequently such as SVM, SVR, and
k-mean; and others are rarely used (boosting, regression tree, random forest, etc.).
Three methods will be probably used in the few next years: SVM, regression trees
and random forests, because the results obtained today with these methods are very
promising. The deep learning, which is a branch of machine learning based on a set
of algorithms that attempt to model high-level abstractions in data by using model
architecture, with complex structures or otherwise, composed of multiple nonlinear
transformations, is certainly the kind of prediction models that will be abundantly
studied in the near future. This research area is very recent and there is not enough
experience, but in the future this kind of methodology may outperform conventional
methods, as it is already the case in other predicting domains (air quality, wind,
economy, etc.). As a consequence, forecasts reached through various methods can
be calculated to satisfy the various needs. A question then arises how they will be
put together? The answer is clearly not trivial because the various resulting forecasts
show differences on many points. Moreover, some of them will be associated with
confidence intervals, which should also be merged.

7.2 SIX HOURS AHEAD PREDICTION AND MORE
Three kinds of improvements should appear in the 10 next years, the first one con-
cerns the MOS [24], the next one the improvement of NWP, and the last one the use
of global irradiation measurement through the world to validate models and maybe
to propose a robust methodology for all sites and horizons.

7.2.1 Model Output Statistics
MOS is an objective weather forecasting technique that consists of determining a
statistical relationship between a measurement and forecast variables by a numerical
model at some projection time(s). It is, in fact, the determination of the “weather-
related” statistics of a numerical model. Lauret et al. [99] propose an improvement