This document summarizes a study on short-term wind power forecasting for a wind farm in complex terrain in China. The study combines micro-scale computational fluid dynamics modeling with artificial neural networks to minimize forecast errors. Testing was performed from March 2012 to November 2012 with forecasts made every 15 minutes up to 46 hours ahead. Results showed the combined approach reduced mean absolute error by 5% and bias by 42% compared to using just the physical modeling alone.
Wind meteodyn WT cfd micro scale modeling combined statistical learning for short term wind power forecasting
1. We consider here a real wind farm in China with complex terrain and weather
regime (snow fall, highly variable wind regime)
Period of study
Learning : 06/2010 to 02/2012
Testing : 03/2012 to 11/2012
Forecast Horizon +22h to +46h
Forecast steps 15 min - Runs 4/day
Input variables
NWP : wind speed, direction
stability, temp., density, pressure
Park production
Short-term wind power forecasting is now considered as a mature field. It has been
widely put into operation within the past ten years. Meteodyn with over a decade of
experience in wind engineering has contributed to this spread with tens of wind
farm equipped with forecast products around the world.
Our next-generation short-term forecasting solution has been designed to makes
the most of both a tailored micro-scale CFD modeling and advanced statistical
learning based on neural networks.
We present here both our innovative strategy to reduce error an evaluation of the
overall performance from real wind farm in complex terrain.
Abstract
Micro-Scale Modeling combined with Statistical Learning
for Short-Term Wind Power Forecasting
Jérémie JUBAN – Jean-Claude Houbart – María BULLIDO GARCIA – Didier DELAUNAY
METEODYN, 14 Bd. Winston Churchill, Nantes, France
PO. IDPO. ID
288288
Case Study
Objectives
Methods
It is widely admitted [1] that an optimal combination of both physical and statistical
modeling allows to reach the highest forecast performance.
Statistical learning based on past measures allows to reach high performance
forecasts by explicitly minimizing average scores such as MAE and RMSE.
However, it is based on the hypothesis that future will remain similar to the past or
at least the recent past (stationary hypothesis). On the other hand, physical
modeling allows to predict environment changes such as maintenance,
repowering, forest cut.
Meteodyn Forecast already benefits from an advanced physical modeling based
on our expertise with Meteodyn WT. Our software features tailored mesoscale
NWP inputs, state-of-the art micro-CFD modeling including atmospheric stability,
forest model, turbines wakes, advanced power curve modeling (WTG).
The objective here is to minimize the prediction error by introducing automatic
error correction while keeping the advantage of our full physical modeling.
Results
12
12.5
MeanAbsoluteError(%Nom
WT
WT+ANN
Mean Absolute Error on testing set - Overall Relative Improvement 5 %
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
Bias(%NomPow)
Look Ahead Time (h)
WT
WT+ANN
Bias on testing set - Overall Relative Improvement 42 %
Conclusions
EWEA 2013, Vienna, Austria: Europe’’’’s Premier Wind Energy Event
We consider here a “black box” statistical correction based on park production
measurements. The “black-box” learning model used here are Artificial Neural
networks.
Artificial Neural Networks (ANN) date backs to early work of P. Werbos, D.E.
Rumelhart, G.E. Hinton and R.J. Williams in the 70’s. ANN are now widely used as
“black box” models in various field [2]. A fully-connected ANN architecture can be
defined by two quantities: the number of layer and the number of neurons per
layer.
For example : is a 2-3-2-1 Artificial Neural Network.
There exist a vast number of possible network, with highly variable performances.
A too simple neural network is not able to fully correct. A two complicated network
will over-fit the input data. One classical way to select among possible networks
architectures is to use Genetic Algoritms [3]. Based on that we have developed a
tailored architecture optimization comprising input variable selection .
Mesoscale NWP
Measurements
Physical Modeling Neural Network
Genetic Algorithm example: 1) start with 3 initial networks; 2) compute their performance 3)
Retain only the best networks 4) Cross the best networks to get possible better betworks.
It has been widely recognize that an optimal combination of statistical and physical
modeling is central to high performance forecasting [1]. However, combining both
in an optimal way remains a difficult task. Based on our expertise in physical
micro-CFD modeling coupled with an advanced statistical learning correction, we
have demonstrated our ability to reach high performance forecasting, with an error
(RMSE) reduce to a15%-16% bound for the next day (+22hto +46h) for highly
complex terrain and weather regimes. We confirmed here that introducing
advanced statistical learning leads to significant improvement over a pure (even
advanced) physical approach (16% relative improvement).
Increasing Model Complexity
9.5
10
10.5
11
11.5
22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
MeanAbsoluteError(%Nom
Pow)
Look Ahead Time (h)
9
11
13
15
17
19
21
22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
RootMeanSquareError(%
NomPow)
Look Ahead Time (h)
WT
WT+ANN
Root Mean Square Error on testing set - Overall Relative Improvement 16 %
Reference
[1] : Giebel G., Brownsword R., Kariniotakis G., Denhard M., Draxl C. The State-Of-The-Art in Short-Term Prediction of
Wind Power A Literature Overview, 2nd Edition. Project report for the Anemos.plus and SafeWind projects., Risø,
Roskilde, Denmark, 2011
[2] : Smith, Murray, Neural Networks for Statistical Modeling, Van Nostrand Reinhold, 1993.
[3] : D. Whitley, "Applying Genetic Algorithms to Neural Network Problems," International Neural Network,1988