A review of state-of-the-art machine learning methods for improving probabilistic weather forecasts

27 november 2020

KNMI scientists, Dr Maurice Schmeits and Dr Kirien Whan, have contributed to a review of the latest methods to improve weather forecasts using statistical and machine learning methods. The review was led by Dr St├ęphane Vannitsem from the Belgium Royal Meteorological Institute (RMI) and had contributions from scientists working on statistical post-processing at 11 national weather services (NWSs) in Europe, as well as Karlsruhe Institute of Technology, and the European Center for Medium-Range Weather Forecasts.

The review was conducted as part of the EUMETNET activity on post-processing. It was written after scientists met in Brussels in December 2019 (Figure 1) to review the post-processing and blending activities of European NWSs and to discuss new ideas emerging in the field. 

The review paper summarizes the main ways in which we use machine learning methods to improve forecasts, from theoretical developments to operational applications, with a focus on the current challenges and potential avenues in the field.

There are many probabilistic post-processing methods that are currently available. These can be seen below in Figure 2. These methods range in their flexibility and complexity, and differ in the assumptions that must be made. The techniques range in complexity from simple bias-corrections to very sophisticated distribution-adjusting algorithms that incorporate correlations. 

At KNMI we have extensive experience using machine learning methods, such as the tree-based “Quantile Regression Forests” (QRF). We use these methods to understand relationships between variables, and to improve forecasts of precipitation, wind speed, solar radiation, and temperature. QRF is a very flexible method that is reasonably easy to implement, is able to capture non-linear relationships, and does not make assumptions about the distribution.

The review article discussed challenges that face the statistical post-processing community, such as the preservation of correlations in space, time and between variables, how to cope with changes in the dynamical forecast model, and the transfer of knowledge from research to operations.

See the article for more information: Statistical Postprocessing for Weather Forecasts – Review, Challenges and Avenues in a Big Data World

A photo from the workshop dinner
Figure 1. Scientists from 11 national weather services met to discuss statistical post-processing and blending in December 2019 (photo credit: Jonathan Demaeyer and Lesley De Cruz, RMI).
A figure outlining post-processing methods
Figure 2. Overview of probabilistic post-processing methods, showing the distributional assumptions (colors), the ease of implementation and the flexibility of each method (from Vannitsem et al, 2021).