Intelligent estimation of wind farm performance with direct and indirect 'point' forecasting approaches integrating several NWP models

Reliable wind power forecasting is essential for profitably trading wind energy in the electricity market and efficiently integrating wind-generated electricity into the power grids. In this paper, we propose short-and medium-term wind power forecasting systems targeted to the Nordic energy market, which integrate inputs on the wind flow conditions from three numerical weather prediction sources. A point forecasting scheme is adopted, which forecasts the power at the individual turbine level. Both direct and indirect forecasting ap-proaches are considered and compared. An automated machine-learning pipeline, built and optimized using genetic programming, is implemented for developing the proposed forecasting models. The turbine level power forecasts using different approaches are then combined into a single forecast using a weighting method based on recent forecast errors. These are then aggregated for the wind farm level power estimates. The proposed fore-casting schemes are implemented with data from a Norwegian wind farm. We found that in both the direct and indirect forecasting approaches, the forecasting errors could be reduced between 8% and 22%, while inputs from several NWP sources are used together. The wind downscaling model, which is used in the indirect forecasting approach, could significantly contribute to the model's accuracy. The performance of both the direct and indirect forecasting schemes is comparable for the studied wind farm.

Automated summary

This paper presents short- and medium-term wind power forecasting systems for the Nordic energy market. Multiple numerical weather prediction sources are integrated to predict power at the individual turbine level. Both direct and indirect forecasting approaches are considered and compared, using an automated machine-learning pipeline. The proposed forecasting schemes reduce forecasting errors by 8% to 22% when using inputs from multiple NWP sources, and the wind downscaling model significantly improves accuracy.