Ensemble climate forecast post-processing with a versatile probabilistic neural community method
Authors: Peter Mlakar, Janko Merše, Jana Faganeli Pucer
Summary: Ensemble forecast post-processing is a needed step in producing correct probabilistic forecasts. Typical post-processing strategies function by estimating the parameters of a parametric distribution, regularly on a per-location or per-lead-time foundation. We suggest a novel, neural network-based technique, which produces forecasts for all areas and lead occasions, collectively. To loosen up the distributional assumption of many post-processing strategies, our method incorporates normalizing flows as versatile parametric distribution estimators. This permits us to mannequin various forecast distributions in a mathematically precise means. We display the effectiveness of our technique within the context of the EUPPBench benchmark, the place we conduct temperature forecast post-processing for stations in a sub-region of western Europe. We present that our novel technique displays state-of-the-art efficiency on the benchmark, outclassing our earlier, well-performing entry. Moreover, by offering an in depth comparability of three variants of our novel post-processing technique, we elucidate the explanation why our technique outperforms per-lead-time-based approaches and approaches with distributional assumption
