Statistical Postprocessing of Ensemble Forecasts of Wind

Abstract

This thesis considers methods and models for postprocessing ensemble forecasts of wind. Based on Bayesian model averaging (BMA), several different extensions are proposed and tested. Firstly, historical observations of wind speed are included in the model as forecasts, both as a climatology and as an ensemble. Secondly, an extension to the BMA in which thin plate regression splines over both forecast wind speed and forecast wind direction are used in the modelling of the expectation of the predictive probability density functions (PDFs) is tested. Each method is assessed mainly using the continuous rank probability score (CRPS), but certain aspects of the forecasts, such as their performance for stronger winds, are assessed using the Brier score and the quantile score. We identify a shortcoming of the BMA involving bias in the forecasting of stronger winds, and an amendment to the method is proposed. This extension is shown to produce better forecasts and goes a long way towards solving the problem with bias in the forecasts of stronger wind.