Machine learning seminar series - From research to applications – Examples of operational ensemble post-processing using machine learning

11:30 BST | October 1, 2020


Zied Ben Bouallegue (ECMWF)


Maxime Taillardat received a M.Sc. Degree in computer science from the National Institute of Electrical engineering, Electronics, Computer science, Fluid mechanics & Telecommunications and Networks, Toulouse, France, a M.Sc. Degree in meteorology, statistics & machine learning from the National   Meteorology   School,   Toulouse,   France,   and   the   Ph.D.   degree   in   meteorology, oceanography and environmental science from Université Paris-Saclay, Versailles, France. He works in the Statistical Forecasting and Verification team in Météo-France, the French Weather Service. He is also affiliated to the National Centre for Meteorological Research – UMR 3589, Toulouse, France. His research interests include among others the use of machine learning algorithms in weather forecasting, especially for the post-processing of numerical weather prediction models, decision sciences, and the verification of ensemble forecasts for extreme events. Since 2018, he is one of the conveners of the session ''Advances in statistical post-processing for deterministic and ensemble forecasts'' at the European Geosciences Union General Assembly.


Statistical post-processing of ensemble forecasts, from simple linear regressions to more sophisticated techniques, is now a well-known procedure in order to correct biased and poorly dispersed ensemble weather predictions. However, practical applications in National Weather Services is still in its infancy compared to deterministic post-processing. We present two different applications of ensemble post-processing using machine learning at an industrial scale. The first is a station-based post-processing of surface temperature in a medium resolution ensemble system. The second is a gridded post-processing of hourly rainfall amounts in a high-resolution ensemble prediction system. The techniques used rely on quantile regression forests (QRF) and ensemble copula coupling (ECC), chosen for their robustness and simplicity of training whatever the variable subject to calibration.

Moreover, some variants of classical techniques used such as QRF or ECC have been developed in order to adjust to operational constraints. A forecast anomaly-based QRF is used for temperature for a better prediction of cold and heat waves. A variant of ECC for hourly rainfall is built, accounting for more realistic longer rainfall accumulations. It is shown that forecast quality as well as forecast value is improved compared to the raw ensemble, which is critical in the context of forecast automation. At last, comments about model size and computation time will be done.