Physics informed machine learning for wind speed prediction

The ability to predict wind is crucial for both energy production and weather forecasting. Mechanistic models that form the basis of traditional forecasting perform poorly near ground. Here we take an alternative data-driven approach based on supervised learning. We analyze massive datasets measured from anemometers located at 10 m height in 32 locations central north-west Italy. train learning algorithms using past history its value future horizons. Using data single horizons, compare systematically several where vary input/output variables, memory linear vs non-linear model. then performance best across all find optimal design as well change with location. demonstrate presence a diurnal cycle provides rationale understand this variation. conclude systematic comparison state art algorithms. When focusing publicly available datasets, our algorithm improves 0.3 m/s average. In aggregate, these comparisons show that, when model accurately designed, shallow are competitive deep architectures.