Postprocessing ensemble weather predictions to correct systematic errors has become a standard practice in research and operations. However, only few recent studies have focused on ensemble postprocessing of wind gust forecasts, despite its importance for severe weather warnings. Here, we provide a comprehensive review and systematic comparison of eight statistical and machine learning methods for probabilistic wind gust forecasting via ensemble postprocessing, that can be divided in three groups: State of the art postprocessing techniques from statistics (ensemble model output statistics (EMOS), member-by-member postprocessing, isotonic distributional regression), established machine learning methods (gradient-boosting extended EMOS, quantile regression forests) and neural network-based approaches (distributional regression network, Bernstein quantile network, histogram estimation network). The methods are systematically compared using six years of data from a high-resolution, convection-permitting ensemble prediction system that was run operationally at the German weather service, and hourly observations at 175 surface weather stations in Germany. While all postprocessing methods yield calibrated forecasts and are able to correct the systematic errors of the raw ensemble predictions, incorporating information from additional meteorological predictor variables beyond wind gusts leads to significant improvements in forecast skill. In particular, we propose a flexible framework of locally adaptive neural networks with different probabilistic forecast types as output, which not only significantly outperform all benchmark postprocessing methods but also learn physically consistent relations associated with the diurnal cycle, especially the evening transition of the planetary boundary layer.

No abstract

Abstract. Strong winds associated with extratropical cyclones are one of the most dangerous natural hazards in Europe. These high winds are mostly associated with five mesoscale dynamical features: the warm (conveyor belt) jet (WJ); the cold (conveyor belt) jet (CJ); cold frontal convection (CFC); strong cold-sector winds (CS); and, at least in some storms, the sting jet (SJ). The timing within the cyclone's life cycle, the location relative to the cyclone core and some further characteristics differ between these features and, hence, likely also the associated forecast errors. Here, we present a novel objective identification approach for these high-wind features using a probabilistic random forest (RF) based on each feature’s most important characteristics in near-surface wind, rainfall, pressure and temperature evolution. As the CJ and SJ are difficult to distinguish in near-surface observations alone, these two features are considered together here. A strength of the identification method is that it works flexibly and is independent of local characteristics and horizontal gradients; thus, it can be applied to irregularly spaced surface observations and to gridded analyses and forecasts of different resolution in a consistent way. As a reference for the RF, we subjectively identify the four storm features (WJ, CS, CFC, and CJ and SJ) in 12 winter storm cases between 2015 and 2020 in both hourly surface observations and high-resolution reanalyses of the German Consortium for Small-scale Modeling (COSMO) model over Europe, using an interactive data analysis and visualisation tool. The RF is then trained on station observations only. The RF learns physically consistent relations and reveals the mean sea level pressure (tendency), potential temperature, precipitation amount and wind direction to be most important for the distinction between the features. From the RF, we get probabilities of each feature occurring at the single stations, which can be interpolated into areal information using Kriging. The results show a reliable identification for all features, especially for the WJ and CFC. We find difficulties in the distinction of the CJ and CS in extreme cases, as the features have rather similar meteorological characteristics. Mostly consistent results in observations and reanalysis data suggest that the novel approach can be applied to other data sets without the need for adaptation. Our new software RAMEFI (RAndom-forest-based MEsoscale wind Feature Identification) is made publicly available for straightforward use by the atmospheric community and enables a wide range of applications, such as working towards a climatology of these features for multi-decadal time periods (see Part 2 of this paper; Eisenstein et al., 2022d), analysing forecast errors in high-resolution COSMO ensemble forecasts and developing feature-dependent post-processing procedures.

While previous research on sub-seasonal tropical cyclone (TC) occurrence has mostly focused on either the validation of numerical weather prediction (NWP) models, or the development of statistical models trained on past data, the present study combines both approaches to a statistical–dynamical model for probabilistic forecasts in the North Atlantic basin. Although state-of-the-art NWP models have been shown to lack predictive skill with respect to sub-seasonal weekly TC occurrence, they may predict the environmental conditions sufficiently well to generate predictors for a statistical model. Therefore, an extensive predictor set was generated, including predictor groups representing the climatological seasonal cycle (CSC), oceanic, and tropical conditions, tropical wave modes, as well as extratropical influences, respectively. The developed hybrid forecast model is systematically validated for the Gulf of Mexico and Central Main Development Region (MDR) for lead times up to five weeks. Moreover, its performance is compared against a statistical approach trained on past data, as well as against different climatological and NWP benchmarks. For sub-seasonal lead times, the CSC models are found to outperform the NWP models, which quickly loose skill within the first two forecast weeks, even in case of recalibration. The statistical models trained on past data increase skill over the CSC models, whereas even greater improvements in skill are gained by the hybrid approach out to week five. The vast majority of the additional sub-seasonal skill in the hybrid model, relative to the CSC model, could be attributed to the tropical (oceanic) conditions in the Gulf of Mexico (Central MDR).

Abstract. Strong winds associated with extratropical cyclones are one of the most dangerous natural hazards in Europe. These high winds are mostly associated with five mesoscale dynamical features, namely the warm (conveyor belt) jet (WJ), the cold (conveyor belt) jet (CJ), cold-frontal convective gusts (CFC), strong cold sector winds (CS) and – at least in some storms - the sting jet (SJ). The timing within the cyclone’s lifecycle, the location relative to the cyclone core and some further characteristics differ between these features and hence likely also the associated forecast errors. Here we present a novel objective identification approach for these high-wind features using a probabilistic random forest (RF) based on each feature’s most important characteristics in near-surface wind, rainfall, pressure and temperature evolution. As CJ and SJ are difficult to distinguish in near-surface observations alone, these two features are considered together here. A strength of the identification method is that it works flexibly and independent of spatial dependencies and gradients, such that it can be applied to irregularly spaced surface observations and to gridded analyses and forecasts of different resolution in a consistent way. As a reference for the RF, we subjectively identify the four storm features (WJ, CS, CFC, CJ+SJ) in 12 winter storm cases between 2015 and 2020 in both hourly surface observations and high-resolution reanalyses of the German COSMO model over Europe, using an interactive data analysis and visualisation tool. The RF is then trained on station observations only. The RF learns physically consistent relations and reveals mean sea-level pressure (tendency), potential temperature, precipitation amount and wind direction to be most important for the distinction between the features. From the RF we get probabilities of each feature occurring at the single stations, which can be interpolated into areal information using Kriging. The results show a reliable identification for all features, especially for WJ and CFC. We find difficulties in the distinction of CJ and CS in extreme cases, as the features have rather similar meteorological characteristics. Mostly consistent results in observations and reanalysis data suggest that the novel approach can be applied to other data sets without a need of adaptation. Our new method RAMEFI (RAndom-forest based MEsoscale wind-Feature Identification) is made publicly available for straightforward use by the atmospheric community, and enables a wide range of applications, e. g., towards a climatology of these features for multi-decadal time periods (see Part II), analysing forecast errors in high-resolution COSMO ensemble forecasts and to develop feature dependent postprocessing procedures.