A1337
Title: Probabilistic post-processing of wind speed forecasts with explicit modeling of time dependencies
Authors: Katharina Klein - Utrecht University (Netherlands) [presenting]
Sjoerd Dirksen - Utrecht University (Netherlands)
Maurice Schmeits - The Royal Netherlands Meteorological Institute (Netherlands)
Kirien Whan - KNMI (Netherlands)
Abstract: Post-processing NWP forecast data is essential for removing biases and obtaining better-calibrated probabilistic forecasts. Post-processing multiple lead times simultaneously is particularly challenging due to the inherent temporal dependencies, which classical approaches deal with by using a two-step procedure: lead times are first processed independently, and (empirical) copula methods are subsequently applied to restore dependencies. The proposed ARMOS model, a generalization of the widely used ensemble model output statistics (EMOS) for estimating marginal distributions, can circumvent the second step by incorporating temporal dependencies explicitly. It exploits the autoregressive property of forecast errors and yields a multivariate probability distribution for the given weather variable. The ARMOS model is applied to deterministic forecasts from KNMI's Harmonie-Arome model in order to obtain a multivariate parametric distribution for wind speed forecasts from initialization time to 48 hours ahead. Compared to a state-of-the-art two-step approach (EMOS adapted to deterministic forecasts and paired with Schaake-Shuffle), ARMOS shows similar or better performance on different multivariate and univariate evaluation methods. The model is thus effective in post-processing NWP forecasts for multiple lead times without the need to use empirical copula methods.
