Comparison of linear, Kalman filter and neural network downscaling of wind speeds from numerical weather prediction

Abstract

This paper describes a refinement of wind speed prediction methods in order to enhance their accuracy for wind energy applications. Specifically, techniques used to downscale raw forecasts from numerical weather prediction models are investigated. Wind speed measurements from several surface meteorological stations are used to test the downscaling process. While classical downscaling methods require large sets of historical data in order to be trained, the Kalman filter has the potential to rapidly estimate the bias that needs to be added to the raw forecasts in order to provide the best fit possible to local observations. In this paper, the Kalman filter technique is applied, and its performance is compared with classical linear and simple artificial neural network downscaling methods. It is shown that while the levels of prediction accuracy attainable are similar to classical techniques, the amount of data required to parameterise the Kalman filter is much less than for other techniques.