A0865
Title: Statistical post-processing of weather forecasts using engression
Authors: Sam Allen - ETH Zurich (Switzerland) [presenting]
Xinwei Shen - ETH Zurich (Switzerland)
Johanna Ziegel - ETH Zurich (Switzerland)
Abstract: Numerical weather prediction models produce weather forecasts that exhibit systematic biases. To yield more reliable forecasts, statistical post-processing methods are used to recalibrate the weather model output. While statistical post-processing is now well-established within operational weather forecasting suites, most post-processing methods are univariate and, therefore, do not yield coherent forecasts for multiple weather variables, time points, and spatial locations. This can be remedied by modelling the dependence structure between different variables using copulas. However, such approaches typically either lack the flexibility required to model the complex dependencies in the weather or do not allow the inclusion of additional covariates when modelling these dependencies. The proposal is to statistically post-processing weather forecasts using engression. Engression is a distributional regression technique that combines generative machine learning with pre-additive noise, resulting in a simple yet powerful post-processing framework. Engression can be applied multivariately and also exhibits desirable theoretical properties when extrapolating beyond observed data, allowing accurate forecasts to be made at spatial locations for which data is not available, for example. To demonstrate engression as a statistical post-processing method, it is compared to state-of-the-art machine learning-based methods when post-processing temperature and wind speed forecasts in Germany.
