Medium-range prediction of near-surface meteorology with physical modelling and machine learning
Summary of the presentation
A two-step approach is presented to improve surface weather forecasting (temperature, humidity, and winds). The first step involves scaling deterministic atmospheric forecasts using a 2.5 km grid physical surface model. This method is followed by the application of U-Net models trained on operational analyses at ECCC for the atmosphere (15 km grid) and the surface (2.5 km grid). The relative and combined impact of these two methods will be discussed.
Learning objectives :
Kilometer-scale scaling of deterministic atmospheric forecasts with ECCC's Surface Prediction System (SPS)
Development and use of U-Net convolutional neural networks for medium-range surface forecasting (temperature, humidity, winds)
Speaker
Dr. Stéphane Bélair obtained his PhD in atmospheric and oceanic sciences from McGill University. After postdoctoral projects at Météo-France (1995) and Environment Canada (1996-1997), he was hired as a research scientist at Environment and Climate Change Canada (ECCC). He has worked on the representation of clouds, precipitation, land surface, and boundary layer in numerical weather prediction models. He has also contributed to terrestrial data assimilation, urban modeling, and subkilometer-scale modeling. Today, he is focused on applying machine learning to improve environmental forecasting and analysis.
