Page 20 - North American Clean Energy January February 2015

wind power

Advancing Short-term

Wind Power Forecasting

By combining CFD modeling & statistical learning

By Ćline Bezault

Figure 1. A deterministic approach

Figure 2. A stochastic approach Figure 3. One time step of numerical weather prediction (NWP) Figure 4. SCADA information added to the NWP

After almost three decades of active research, short-term wind power forecasting is now considered a mature ield. It has been

widely and successfully put into operation within the 10 past years. Anticipating the amount of energy that’s to be harvested

is key to load balancing and revenue generation for any productive wind farm, and two main approaches for wind power

forecasting are usually considered in the literature (though are sometimes opposed): Physical and statistical.

Physical models tend to draw data from external numerical weather prediction (NWP) For each axis, one concept generally excludes the other. Intraday (very short-term) is
models, and include mesocale and computational luid dynamics (CFD) modeling; whereas, commonly stochastic with online measurements, while extraday (short-term) is usually
statistical models employ diferent statistical algorithms. hese include grey/black box deterministic based on NWP data.
statistical learning, phase/magnitude correction, and data iltering.
herefore, the aim of the ideal wind forecasting tool is to breakdown these classiications

Over time, however, it’s been widely determined that an optimal combination of physical by proposing a unique model, which merges all these techniques. A short-term forecasting
and statistical approaches are necessary to build a high-performance forecasting system.
solution has, in fact, been designed to take advantages of micro-scale CFD modeling and
Behind the optimal combination, there still resides a wide variety of design options. he advanced statistical learning. In the frame of this model design, various options have

following sheds some light on what performances one should expect from several modeling been considered and evaluated, taking into account model performance and operational
options for combining physics and statistics in wind forecasting. he case studies presented constraints (see Figure 1).
are taken from real wind farms in various climate and terrain conditions.
Deterministic forecast: A physical approach

Usually, wind energy forecasting systems are classiied along several axes:
A deterministic approach starts from a NWP, which predicts the meteorological wind over
• Intraday/Extraday, or Short-term/Very short-term. his mode looks at the time
and around a wind farm area. hen, a computational luid dynamics (CFD model) tool
ahead (i.e. horizon), from 0 to several hours for very short-term (intraday), and from
makes a micro-downscaling, so as to provide the future wind characteristics at the exact
some hours to some days for short-term (extraday);
location of each turbine.

• Deterministic/Stochastic. hese two opposite concepts use diferent scientiic tools—
As a result of the measured power curves of the turbines, the air density provided by the
physics and mechanics for a deterministic approach, as well as statistical tools and
numerical weather prediction and the planning maintenance (provided by the user), the
machine learning for stochastic approach;
inal power output can be calculated. his step includes the interaction between the wind
• Numerical weather prediction/Online data. here are basically two sources of data
turbines, throughout the Jensen wake model. (Note: the Jensen model is, currently, one

to perform wind power forecasting. he irst one includes meteorological predictions, of the widely used wake models as it’s not only simple and easy to code, but also performs
based on global measurements and advanced numerical computations; the second is well on predicting wake loss.)
online measurements, such as for instance, from supervisory control and data acquisition
(SCADA) systems.

20 JANUARY/FEBRUARY 2015 nacleanenergy.com