Posters

Abstract

It is important to rely on a wind power forecasting model in order to reduce energy trading risks and to plan maintenance strategically, reducing revenue losses. In this work a comparison between two different approaches to create a power forecasting, downscaling from meteorological forecasts, using machine learning are presented: the first strategy consists in creating one regression model for the whole wind farm and, the second one consists in one regression downscaling model for each wind turbine. To develop these models (1) first it is necessary to calculate historical potential power for each wind turbine at each timestamp, identifying sub performance and calculating energy losses; (2) then collect historical meteorological wind forecasts for the wind farm location; (3) using these data and a parallel computing strategy on google cloud platform, neural network models were trained for each wind turbine and for the wind farm average; (5) in a hold out dataset, predicted values were compared to the operational potential power and the metrics were compared for the two strategies in different wind farms. For a smaller wind farm there were no significant differences between the two strategies, however in larger wind farms, constructing one model for each wind turbine (second strategy ) leads to lower errors.