Machine learning forecasts of Scandinavian numerical weather prediction wind model residuals with control theory for wind energy

Abstract

The quality of wind data from the numerical weather prediction significantly influences the accuracy of wind power forecasting systems for wind parks. Therefore, an in-depth investigation of these wind data themselves is essential to improve wind power generation efficiency and maintain grid reliability. This paper proposes a novel framework based on machine learning for concurrently analyzing and forecasting predictive errors, called residuals, of wind speed and direction from a numerical weather prediction model versus measurements over a while. The performance of the framework is testified by a wind farm inside the Arctic. It is demonstrated that the residuals still contain significant meteorological information and can be effectively predicted with machine learning and the linear autoregression works well for multi-timesteps predictions of overall, East–West,​ and North–South wind speeds residuals by comparing the four forecast learning algorithms’ performance. The predictions may be applied to correct the NWP wind model, making quality feedback improvements for inputs for wind power forecasting systems.

Keywords

• Autoregressive forecast
• Machine learning
• Residual analysis
• Statistical inference
• Wind energy
• Feedback