Postprocessing multi-model ensemble temperature forecasts using Distributional Regression Networks
Enric Casellas Masana, Josep Ramon Miró Cubells, and Jordi Moré Pratdesaba
Introduction
This work presents a comparative analysis of postprocessing techniques for multi-model weather ensemble forecasts (Poor Man's Ensemble).
We explore the application of Ensemble Model Output Statistics (EMOS) using the IMPROVER framework and also two Distributional Regression Network (DRN) approaches to postprocess hourly temperature forecasts.
Go to Introduction ↗Data
27 months of air temperature hourly data
10 member multi-model ensemble
186 automatic weather stations as ground truth
Methodologies
Distributional Regression Networks (DRN), based on Rasp and Lerch (2018), are implemented using TensorFlow. Either the mean and standard deviation of the ensemble or its individual members are utilized in the implementation.
Conclusions
Any postprocessing method improves upon the raw PME forecast.
DRN-Members, which uses PME members as input features, achieves the lowest mean CRPS and the flattest rank histogram for 0-48 hours
IMPROVER performs better than DRN-Mean in CRPS but has poorer rank histograms. DRN-Mean shows a slight positive bias, while IMPROVER displays some underdispersion.