Day-ahead probabilistic wind power forecasting based on ranking and combining NWPs

Dispatching energy in transmission and distribution networks and bidding on electricity markets require probabilistic wind power forecasts available several hours before the actual occurrence. The volatility of the wind over large time horizons complicates the generation of skilled, reliable wind power forecasts. Exploiting numeric weather predictions (NWPs) is generally considered mandatory to increase the skill of probabilistic predictions, and forecasts may further be enhanced by adding several spatially distributed predictions. However, feature selection becomes a more complicated and time consuming as the number of NWPs increases. In this paper, we predict the power generated by a wind farm developing a new technique on the basis of ranking and combining spatially distributed NWPs, easing the feature selection and reducing the computational efforts, as well as maintaining high the skill of probabilistic forecasts. Several spatially distributed NWPs, provided for the area surrounding the wind farm, are ranked for each individual generator, and the ranked NWPs are combined to form an ensemble set of predictors for the probabilistic forecasting model. This ensemble is obtained using three different weighted combination approaches. Gradient boosting regression tree models and quantile regression neural networks generate probabilistic wind power forecasts. The proposed methodology is applied for day-ahead wind power forecasting of individual generators and of the entire wind farm. Numerical experiments carried out on an actual wind farm in southern Italy suggest that ranking NWPs may keep the skill of forecasts at high levels even under a simplified, less computationally intensive procedure. The performance is also enhanced up to 1.8% with respect to standard techniques.