A Forecast Tailored to Your Needs

15 Mar 2017

A data story

It may seem obvious, but when it comes to forecasting the wind, you need to know a wind farm’s production for a given horizon. A multitude of input data is required; the first is weather. This data may arrive in different forms, with several suppliers giving multiple resolutions and types, and could be improved with CFD (Computational Fluid Dynamics) models. CFD codes dedicated to wind resource assessment help to downscale the weather data.

Another type of data is the supervisory control and data acquisition (SCADA) produced by the wind turbine itself. This provides a wealth of information that can be processed to extract useful information for production. The last type of data is comprised of the real-time numbers of this same turbine, associated with local measurements on-site (if these are available).

In order for this information to be used in the most efficient way possible and improve your final production forecast, you must aggregate the data into a module of machine learning. A machine learning system can take various forms depending on the application. In the wind energy sector, the most frequently used is the Artificial Neural Network (ANN).

Weather data

There are several global models of meteorological data (ECMWF, GFS, GSM, GASP, Arpege, UM). Moreover, several suppliers today offer their own weather data, either from proprietary modelling or from an improvement on the global models.

It’s very expensive to buy all the meteorological data from all the global models. Does it make sense to choose just one set of data? If so, which one? Each model has its own characteristics, advantages, and disadvantages. A model that works for one farm may not be appropriate for another. On the other hand, if you average all of the different data too quickly, too much information would be lost.

The current trend is to mix the models using a machine learning system. By taking a period large enough for accurate learning, you can adapt this data to any wind farm in any region.

A test was made on five wind farms with one year of weather data. Penalty savings were compared between two suppliers, as well as using a mix of the two. Notably, mixing data resulted in a gain of up to 37% on penalty savings.

Another way to improve weather data is by using a CFD code. Today, you have access to numerous CFD tools dedicated to wind resource assessment. They take into account a better resolution of ground and roughness, and even the thermal stability of a given location. The CFD downscaling is mainly useful in complex or forested areas, but can also aid in forecasting a new or repowered wind farm. Without historical data, the physical model with CFD is the best way to improve these results.

The SCADA data

SCADA is a veritable reservoir of information (stops, production, over and under production, curtailment, failures), which presents the following problems: What to do with all this data? How to use it? Should you keep it all? Do you need to average the availability of the farm? What about stops?

It’s vital to study this data precisely, arrange it, clean it, find relevant indicators, and bring out the history of the machine. Fortunately, new software has developed to process this data optimally.

The real-time data generated by wind turbines and on-site measuring instruments make it possible to readjust measurement and forecasts. Analysing the state of the machine in real time allows readjusting of the forecast, especially on short times (using a persistent approach, in most cases).


The excessive amount of input data (weather, SCADA, real time, etc.) must be used effectively. How can you mix this data? How can you find the best indicator for a given situation? The most commonly used machine learning system in wind energy is the ANN. The neural network is comprised of several nodes meant to mimic the biological neurons of the human brain. It creates links between each input variable and gives them weight. The goal, by studying the past, is to establish a large database that can identify all possibilities. This will ensure the highest production in any given circumstance.

The ANN must be optimized and calibrated according to the purpose of the forecast (maintenance, trading) and effectively adapt to the rules of these practices.

A study of five wind farms with one year of meteorological data found that optimized mixing allows systematic gains between 20% and 45% on the penalties savings.

Predict the big weather deviations

A problem that arises today, despite the learning and despite the improvement and processing of all these data inputs, is that significant deviations can appear and cause strong penalties. How can they be predicted?

Above all, you need a methodology and adapt it to your needs. You can generate reliability criteria for a given prediction. For example, when you have a deviation, what can you relate it to?

Is there a similarity between several suppliers? How is the accuracy of the deciles? When was the data last updated? Is there any confident parameter from the supplier?

The goal is to create an indicator that triggers an alert. The criteria you select are essential; you will need to arbitrate according to several. Ask yourself the following questions: How many false alarms can I accept? How many risky days without warning can I accept?

In conclusion

Regardless of the slew of data at your disposal (the precision of the input weather data, the improvement of this precision by the CFD, better use of SCADA data and real time, the improvement of learning machines) forecasting is not a one-off. You will always have to study each task, each purpose, and each goal in order to successfully tailor a forecast to your needs.


Bertrand Crouzille is the Communication, Support and Training manager for the worldwide company Meteodyn. Meteodyn develops CFD wind modeling, Operations and Maintenance and forecast software and provides climatological and energy studies (temperature, hygrometry, wind, sunshine, snow). Its consulting services focus on renewables, sustainable construction and city planning, and wind safety.

Meteodyn Meteorology and Dynamics |  http://www.meteodyn.com

Volume: 2017 March/April