The life of GLAMEPS

GLAMEPS (Grand Limited-Area Model EPS) was a pan-European multi-model ensemble prediction system (EPS) for reliable short-range forecasts developed by the HIRLAM and ALADIN consortia. The development started in 2005, and the system was run as one of the first time-critical applications at ECMWF from 2011 to 2019 with an upgrade in 2014. The skillful ECMWF staff was crucial for the implementation and production of GLAMEPS. It included a hierarchy of emergency procedures in real time, and GLAMEPS-staff were contact points in cases when manual interventions were necessary. Computer resources were dedicated by HIRLAM and ALADIN countries from their national ECMWF quotas.

Basic concepts

GLAMEPS aimed to provide predictions up to 2–3 days ahead which accounted for initial state and model inaccuracies and addressed risks of high-impact weather. The system that became operational achieved better probabilistic verification than ECMWF’s ensemble forecasts (ENS) for most near- surface weather parameters (see figure for an example).

The system made use of a few different limited-area numerical weather prediction models, which formed the basis of alternative control forecasts as well as different sets of alternative ensemble forecasts. There was an emphasis on accounting for forecast uncertainty originating from processes at the ground surface. In addition to the higher level of spatial detail than in ENS, the multi-model approach proved valuable for improving the forecasts relative to those produced by ENS for the short range.

Pre-operational GLAMEPS

This prototype was run twice daily up to 42 hours for a seven-week winter period. The results were analysed and published scientifically. The 52 members of GLAMEPS_v0 consisted of three control forecasts, four different models, and EuroTEPS, a model with stochastic physics. EuroTEPS was a version of ENS with perturbations spatially targeted for Europe that was run on behalf of HIRLAM and ALADIN with the same resolution as ENS at that time (~55 km). The horizontal mesh width was ~13 km for the LAM models. The HIRLAM model used two parametrizations of deep convection, which also produced two control runs based on parallel data assimilation (3D-Var). The third control run was from EuroTEPS, while ALADIN downscaled EuroTEPS members.

The forecasts by GLAMEPS_v0 scored better than ENS with respect to ensemble calibration, forecast reliability and information content, and potential economic value. The multi-model approach was important for the improved quality. Since replacing EuroTEPS with ENS degraded the forecasts only slightly, it was decided to use ENS operationally as the global model.

GLAMEPS 2011–2013

This first operational version used all 51 members of ENS, either directly or as boundary and initial perturbations, and in addition the high-resolution forecast from ECMWF (HRES). The model domain was increased by 30% with higher horizontal resolution (~11 km), and separate ground surface data assimilation cycles were run for each ensemble member from the HIRLAM and ALADIN-ALARO models. The latter was motivated by the increased potential predictability of weather features forced by the ground surface.

To optimize the timely use of ENS in ECMWF’s own production schedule, hours 06 and 18 UTC were the base times for the 54 h ensemble forecasts, which added up to 54 members. A selection of standard probabilistic maps and EPS-meteograms were provided, along with raw data (see figure for an example).

GLAMEPS 2014–2019

The upgrade in 2014 included increased horizontal resolution (~8 km) and up to 60 h forecasts produced four times per day (00, 06, 12, and 18 UTC). Ensemble members from ENS were no longer part of the multi-model ensemble but provided perturbations of initial and lateral boundary conditions.

The multi-model approach had four sub-ensembles, each with six new members produced six-hourly. The four sub-ensembles used separate model configurations, two of HIRLAM and two of ALADIN-ALARO. Each model configuration also produced one new control every six hours. By combining the new 24 ensemble members with the 24 from the previous base time, 4 control runs and 48 alternative ensemble members were available six-hourly. Stochastic perturbations of physics tendencies were included in the HIRLAM model.

The upgrade enabled probabilistic forecasts with increased spatial resolution, earlier delivery, more frequent updates, and extended forecast range. The added value over corresponding products from ENS was maintained operationally.

Extensive verification with the HARP validation package developed for GLAMEPS revealed that an important contribution to the skill enhancements of GLAMEPS was the combination of models with comparable forecast quality. Lagging of ensemble perturbations also enhanced the skill. Corresponding single model ensembles produced insufficient internal spread, although it was improved by the stochastic physics tendencies.

Conclusion

GLAMEPS production stopped in 2019. The HIRLAM and ALADIN consortia saw the need to pursue even higher resolution EPS for the very short range, appreciating that very small-scale phenomena often are important for the development of extreme weather. The success of GLAMEPS, as well as its limitations, inspired the ongoing operational and experimental work on convection- permitting EPS in sub-European domains, implementing many of the successful system elements developed for GLAMEPS.