Economical microscale predictions of wind over complex terrain from mesoscale simulations using machine learning

The ability to assess detailed wind patterns in real time is increasingly important for a variety of applications, including wind energy generation, urban comfort and environmental health, and drone maneuvering in complex environments. Machine Learning techniques are helping to develop accurate and reliable models for predicting local wind patterns. In this paper, we present a method for obtaining wind predictions with a higher resolution, similar to those from computational fluid dynamics (CFD), from coarser, and therefore less expensive, mesoscale predictions of wind in real weather conditions. This is achieved using supervised learning techniques. Four supervised learning approaches are tested: linear regression (SGD), support vector machine (SVM), k-nearest neighbors (KNn) and random forest (RFR). Among the four tested approaches, SVM slightly outperforms the others, with a mean absolute error of 1.81 m/s for wind speed and 40.6 ∘ for wind direction. KNn however achieves the best results in predicting wind direction. Speedup factors of about 290 are achieved by the model with respect to using CFD.