Probabilistic weather prediction: From ensembles to neural networks
31. January 2018
Nowadays, weather forecasts are often given in the form of forecast ensembles obtained from multiple runs of numerical weather prediction models with varying initial conditions and model physics. Such ensemble predictions typically tend to be biased and lack calibration, and thus require some form of statistical post-processing. Using forecast-observation pairs from the past to identify model errors and adjust the current forecast accordingly, statistical post-processing allows to correctly represent and quantify the associated prediction uncertainty in order to provide calibrated and sharp predictive distributions, a prerequisite for optimal decision making in many applications. This talk will review recent developments and active areas of research in statistical post-processing of ensemble forecasts spanning from classical distributional regression methods to first attempts of using neural network approaches.
This talk was given by Sebastian Lerch. Sebastian is a Postdoc at the Institute for Stochastics of the Karlsruhe Institute of Technology (KIT), and a member of the Computational Statistics group at the Heidelberg Institute for Theoretical Studies (HITS).