URJC and MIT lecturers have presented the results of a study on more accurately estimating and predicting wind speed distribution, using less data than standard models used in the sector.
Undertaken jointly by Professor Alfredo Cuesta-Infante, currently a researcher from the King Juan Carlos University Optimisation and High Performance Computing group (OHPC), and professors Kalyan Veeramachaneni and Una-May O’Reilly from the Computer Science and Artificial Intelligence Laboratory (CSAIL), belonging to Massachusetts Institute of Technology (MIT), the work was presented at this year’s International Joint Conference on Artificial Intelligence (IJCAI’15), held in Buenos Aires. The researchers have designed a probability model to estimate wind resources. The model relates wind distribution in different locations to what exists at the location we wish to estimate or predict.
Many factors need to be taken into account when deciding the location for a wind generation plant, but one of the most important is clearly correctly assessing wind power in that location. Power is related to the wind speed distribution and its direction.
The usual method involves placing a series of anemometers around the location of interest and taking measurements over a set period, which is usually between 8 and 12 months. These measurements are then correlated against historical data collected during the same period in meteorological stations, airports, etc., to assess the suitability of the chosen site.
This approach involves an assumption that is almost never appropriate, that the distribution of the real data is a Gaussian curve, that is, it is supposed the data is normal or has a normal distribution. This is not found to be completely accurate if the data is analysed, however.
The model presented by Professors Veeramachaneni, O’Reilly, and Cuesta-Infante firstly uses distributions capable of modifying the bell shape, called ‘Gaussian copula functions’, which improve the results across the board. All the same, this type of function is not yet capable of correctly estimating the extreme cases (a lot or very little wind), given that, as Dr. Cuesta-Infante from the URJC explains, “the issue is that we continue to have the Gaussian curve in the underlying dependency structure. In other words, it is as if we stuck some decorations to the bell that changes it externally”.
As such, the work adds a second modelling technique that only considers non-Gaussian copulas. With this technique, called Vines, the results are much better, as was to be expected.
To undertake the work, they installed anemometers on the roof of the Museum of Sciences, as a site of interest, in Cambridge (Massachusetts), and data was collected for two years. Additionally, historical data was collected from fourteen airports across New England and New York. The first year was used to learn the models and the second year was used to validate them.
What is most interesting about the results is that with just three months of data collected, using very affordable material – the anemometers installed are cheap and airport data is in the public domain – it is possible to estimate much more precise models of wind speed distribution.

Full Article:
Copula Graphical Models for Wind Resource Estimation

Bibliographic reference:

Kalyan Veeramachaneni, Alfredo Cuesta-Infante, Una-May O’Reilly, Copula Graphical Models for Wind Resource Estimation, Proc. of the 24th Int. Joint Conf. on Artificial Intelligence (IJCAI 2015), pp. 2646 — 2654
Source: URJC