Improving wind power prediction with retraining machine learning algorithms

This paper presents a 48-h prediction methodology for wind power production using a machine learning algorithm and focuses on the optimization of the input dataset. While power-grid operators have to keep the production equal to demand, wind power depends on meteorological conditions. Therefore, the main issue of power-grid operators is to predict the wind production as precisely as possible. Our goal is to improve wind prediction accuracy based on the feedback of the relevant work on the subject. To this purpose, past power production and Numerical Weather Predictions (NWP) are used as input of a Gradient Boosting Tree algorithm. Our approach is to lay emphasis on the input data in order to extract as much knowledge as possible and remove as much error as possible from the source. Therefore, 20 days have been removed and additional values have been calculated to improve the algorithm accuracy. The novelty of our approach is in the constant retraining of the model to give the latest information available contrary to most studies, which use a fixed learning dataset. The idea is to combine the advantage of regression models and machine learning algorithms, which use large datasets, in order to learn the relationship between wind production, NWP, and date and time information. The whole methodology helps the model anticipate error of wind prediction and seasonal trends so that the overall prediction accuracy increases by 17% compared to the persistance approach. Using 9 months of historical values from 2016.12 to 2017.08 and predicting from 2017.09 to 2018.01, a prediction accuracy of 83% has been achieved; 17% better than the persistence model. Result analysis show that more improvement can be achieved with a focus on low wind speed values, an improvement of weather prediction using real local weather measures and a reduction of the forecast horizon. The whole prediction process have been automatized and a visualization web page have been implemented.