The American Meteorological Society (AMS) held between January 6th and 10th 2013 its 93rd annual congress.
This prestigious society promotes development, communication, information and training related to hydrological and oceanic sciences from the meteorological point of view.
The multidisciplinary team composed of Lloyd Treinish, James Cipriani and Anthony Praino from IBM Thomas J. Watson Research Center, and Julia Díaz and José Dorronsoro from IIC, presented the first results of their cooperation under the lecture titled Precision Wind Power Forecasting via Coupling of Turbulent-Scale Atmospheric Modelling with Machine Learning Methods, which analyses the difficulty of making a forecast of meteorological variables related to wind power in areas experimenting frequent variations or ramps in weather conditions. It also puts forward possible solutions based on automatic learning methods.
In particular, the conference presents the first steps carried out in the cooperation project between IIC and IBM Thomas J. Watson Research Center, whose geographical area of application covers five of the seven islands that shape the archipelago of the Canary Islands. The effect of turbulent flow is particularly present in the Canaries, as this is an isolated geographical area with a complex orography that favours the appearance of vortexes as altitude increases.
Current techniques for wind power forecasting, as well as for other renewable energies, base their methodology on the creation of predictive models upon historical information in the form of patterns containing energy production data associated with each meteorological forecast. Even though numerical models such as NWP (Numerical Weather Prediction) provide meteorological forecasts of up to several days ahead with high spatial resolution, it becomes essential to accurately determine potential weather events that could cause abrupt variations in power generation in wind farms. The main reasons for research in fluctuation detection systems are to improve quality in wind power forecasting and to reduce additional costs that adjustment in electricity production requires whenever these events occur.
Turbulent flow detection with Deep Thunder
With the aim of accurately detect spontaneous variations in wind power, IBM Thomas J. Watson Research Center has developed a forecasting system based on physical models called Deep Thunder, its main advantage lying in the fact that it provides forecasts every five minutes, with a one-day horizon and a resolution of 0.006º approx. These frequent and rigorous forecasts guarantee a high level of ramp detection. The system was initially configured by means of a series of numerical experiments carried out during several days in 2010 and 2011 in which strong wind currents occurred in the Canarian archipelago. IIC actions involve interpretation and selection of variables which could be more relevant when issuing wind power forecasts.
Predictive models and simplification of information
The main challenge posed by the high resolution meteorological forecast obtained by Deep Thunder is the huge volume of information that each system output involves. Thanks to the automatic learning techniques integrated in the IIC wind power production forecasting system, called EA2, it is possible to process these large data volumes.
In particular, the EA2 system is capable of optimally generating the information patterns provided by Deep Thunder and of training predictive models according to historical information of meteorology and production, using current techniques such as Support Vector Machines – SVMs.
Integration of EA2 and Deep Thunder systems
Therefore, research activities being carried out in the cooperation project between IIC and IBM Watson Research Center have as their main purpose to reduce the volume of information provided by Deep Thunder, to select the most suitable meteorological variables and to take advantage of their forecast to improve the wind power production predictive models that EA2 issues for the Canary Islands.