Spatially coherent postprocessing of cloud cover forecasts using generative adversarial networks

Despite considerable improvements over the last few decades, numerical weather prediction (NWP) models still tend to exhibit bias and dispersion errors. Statistical postprocessing reduces these errors and allows quantifying predictive uncertainty. However, classical postprocessing approaches such as ensemble model output statistics (EMOS) destroy any physical dependence structure of the NWP raw ensemble forecasts. Ensemble copula coupling (ECC) is a commonly used state-of-the-art method to map the spatio-temporal dependence structure of the raw ensemble to the postprocessed predictive distributions. However, if the variable of interest exhibits many ties, ECC may not be optimal. Here, the variable investigated is hourly cloud cover over Switzerland. The climatological distribution of cloud cover shows considerable point masses at both zero and one, hence ties are a major issue when it comes to applying ECC. 

We compare a variant of ECC, which is tailored to variables with many ties, applied to postprocessed forecast ensembles obtained by either EMOS or a dense neural network (dense NN) with postprocessed scenarios generated by a conditional generative adversarial network (cGAN). In particular, cGANs are appealing as they directly generate maps of postprocessed cloud cover forecast scenarios without the need of any dependence template. We trained the postprocessing models for COSMO-E and ECMWF IFS raw ensemble forecasts against hourly EUMETSAT CM SAF satellite data with a spatial resolution of around 2 km over Switzerland. For all the approaches, EMOS, dense NN, and cGANs, basic setups with a minimal set of raw ensemble predictors already allowed us to obtain a significantly better univariate performance (in terms of continuous ranked probability score) than the raw NWP ensembles. We present and discuss the advantages and drawbacks of EMOS+ECC, dense NN+ECC, and cGANs with respect to both univariate forecast skill and the ability to produce realistic cloud cover forecast scenario maps.