Wind Power Forecasting for the Danish Transmission System Operator Using Machine Learning

Abstract

A power grid with increasing wind power and decreasing capacity of conventional power plants induces challenges in the balancing of the power grid. The cost of purchasing reserves in Denmark has increased rapidly over the last five years. One solution to decrease the reserve cost is by introducing new market players to the markets, e.g. wind turbines. Today the wind turbines are excluded from the markets due to low availability. By developing a wind power forecasting model, the availability of the wind at varying wind speeds can be evaluated. A time series neural network with three hidden neurons and two delays are developed. It was found that the highest performance was reached by applying PCA and by using the training algorithm scaled conjugate gradient. The optimal network resulted in an R2-value at 0.990 and MSE at 33895, when testing the model on unseen data. Using the developed model, the availability of wind power was estimated. Limits of the reserve purchase were set at varying wind speeds. The highest purchase was at wind speeds above 20 m/s, where 92% of the predicted power is available with a security of 95%. As the wind speed decreases the purchase decreases as well. The model showed the poorest predictions at wind speeds between 0-5 m/s.