Evaluation of orography and roughness model inputs and deep neural network regression for wind speed predictions

Abstract

Abstract Many wind energy projects start with resource assessments based on outputs from numerical models followed by meteorological campaigns. These models are imperfect and have a substantial input parameter space. It is difficult to discern how the inputs affect the simulation results. Surface roughness is generally crudely represented in mesoscale models. In this work, we use the Weather Research and Forecast model to simulate winds at a meteorological mast in Mexico with wind observations at 80 m. The model sensitivity to changes in surface roughness is contrasted with permissively perturbed orography fields to gauge the observed changes' relative importance. Changes in surface roughness affect the root mean square error of the simulated 80‐m winds compared to the observations, in the same order of magnitude as orography perturbations. Using a roughness field derived from a synthetic aperture radar improved the wind speed bias. Wind speed predictions using high‐resolution (185 m) simulations sometimes showed excellent agreement with the observations, but there were several instances when medium‐resolution (1.6 km) simulations performed better. A wind speed time series with high temporal (10 min) resolution from a low spatial (6 km) resolution simulation was used to train a deep neural network regression (DNNR) model. The trained DNNR reduced the wind speed error the most compared to any of the other simulations performed in this study. The second‐best model performance was obtained using the radar roughness‐derived field. The results of the simulations at medium resolution with perturbed orography were very similar to those using different roughness fields inputs.

Keywords

• machine learning
• mesoscale model
• resource characterisation
• wind speed prediction