Power companies observe proposed sites for wind farms for up to a year to determine their power-generating capacities. Researchers from MIT have come up with an alternative evaluation method, needing just three months' worth of data.
The new method will be presented later in the month at the International Joint Conference on Artificial Intelligence, providing a new statistical technique that offers just-as-accurate predictions for less data. This saves power companies a lot of resources, especially in evaluating offshore sites because maintaining measurement stations can be costly.
When the researchers talked to those in the wind industry, they found that a simple mechanism is being used to estimate wind capacity at a certain site. Kalyan Veeramachaneni from the MIT Computer Science and Artificial Intelligence Laboratory further explained that the industry's standard practice was to assume that data on wind speed follows a Gaussian distribution.
To assess a site, a consultant for a power company will look for correlations between measurements of wind speed made at the site and those at a weather station close by for the same period. Using the correlations they find, the consultant then adjusts historical data provided to come up with an approximate wind speed for the site.
In statistics, using correlations is known as joint distribution, meaning it represents probabilities not only for a certain site but those as well from coincidences that occur for a particular measurement. It's typical for consultants in the wind industry to characterize joint distributions as Gaussians, noted the researchers.
Called a copula graphical model, the model developed by MIT researchers factors in data from several weather stations. However, its main benefit is that it is not limited to probabilities in Gaussian distributions. It can also use different kinds of measurements from different sites, combining data in different ways and spot non-linear connections between sets of data.
Using the copula graphical models tweaked with three months of historical data for a particular wind farm site, the researchers were able to predict at what speeds the winds will come in the next two years at thrice the accuracy of the standard evaluation method's prediction utilizing eight months of data.
Experts of the copula statistical technique laud the work Veeramachaneni and colleagues did, saying there are significant applications in the real world that could benefit from their discovery, and not just the wind industry.
Aside from Veeramachaneni, Una-May O'Reilly and Alfredo Cuesta-Infante also contributed to the research.
Photo: Kevin Dooley | Flickr