Machine learning and Analog Ensemble techniques for temporal extrapolation of wind resource uncertainty

Abstract:

To accurately plan and manage wind farms, not only the average wind resource at the site of interest needs to be assessed, but also the uncertainty connected to this estimate. The quantification of wind resource uncertainty is especially important for offshore wind energy, given the challenges connected to the direct observation of the wind flow offshore. Numerical weather prediction models represent a valuable way to characterize the wind resource offshore. The uncertainty associated to modeled wind speed can be estimated by running multiple ensemble members, obtained by perturbing initial conditions, external forcings, or numerical schemes. However, creating such a numerical ensemble of long-term wind resource data over a large region represents a challenge given the substantial computational capacity it requires.

Here, we quantify uncertainty in hub height wind speed in a region covering the proposed offshore wind energy lease areas in California. We run the Weather Research and Forecasting (WRF) model version 4.1.2 with 16 ensemble members for a single year, while a single WRF run is performed for the remaining 19 years of the period of interest. We apply and compare two different approaches to derive uncertainty estimates without the need of running numerical ensembles over the long period of interest. As a baseline, we use the Analog Ensemble technique to generate uncertainty estimates over the 20-year period. We contrast this approach against a novel approach in which a gradient boosting machine learning model is trained on the single year of WRF ensembles to extrapolate wind speed uncertainty to the remaining 19 years of single model run.

Wind speed uncertainty predicted by the machine learning model has a better agreement with the WRF ensemble spread. On the other hand, Analog Ensemble generally predicts lower uncertainty. Notably, the uncertainty predicted by both methods increases towards the coast (Figure 1). For one of the proposed wind energy lease areas analyzed, hourly uncertainty in hub height wind speed is predicted to be between 5 and 10%. In our presentation, we will examine cases where the predictions from the two approaches differ the most, and evaluate the consequences for offshore wind energy development.