Improving and enhancing NWP based wind power forecasts under Norwegian conditions

Abstract

This thesis studies methods for improving and enhancing NWP based wind power forecasts for cases from the Western coast of Norway. This is an area with excellent wind conditions, but where the installed wind power capacity at present is limited. The area is characterized by a rugged coastline and complex terrain, which have earlier been shown to lead to high wind power forecast errors. The overall aim of the thesis is to study how this kind of conditions influence wind power forecast errors and to investigate how wind power forecast models can be made more resistant to the challenges these conditions pose. The data basis for the thesis consists of wind observations and NWP wind forecast for 43 sites, all covering the period from January 1st 2009 to December 17th 2011. The observations and forecasts are transformed into synthetic wind power forecasts and observations by the use of a logarithmic height profile and a generic power curve. Different methods of reducing the wind power forecast errors of the single sites and groups of sites are tested. When applied to “unseen” forecasts (i.e. independent test data) the simpler models tend to out-compete more complicated models of the same kind. This is caused by a combination of noisy data and sparse data for certain wind speeds and wind directions leading to models being easily over-fitted. Still, using a regression model based on information on spatial and temporal dependencies of the forecast errors of groups of sites, a reduction in the group forecast error of 49 % is obtained. The group point forecasts are expanded into probabilistic forecasts using the post-processing method Bayesian Model Averaging (BMA). It is shown how BMA can be used to produce probabilistic forecasts for the lumped power output of groups of sites from the point forecasts for the single-site group members and historical observations and forecasts from a training period. Some ideas for further development of the method for wind power forecasting are presented. Last, the issue of wind power ramps – large sudden changes in the wind power production – is addressed. Using a simple method to forecast ramps from wind power forecasts it is found that the methods earlier used to reduce the wind power forecast errors for groups of sites also lead to an increased predictability of wind power ramps. Ramp forecasts are also made using the classification method Random Forests. The method is found to have some very desirable properties, but the current implementation of the method also has some serious problems that need to be solved before the method is a good option for wind power ramp forecasting.