This year’s Using ECMWF’s Forecasts (UEF2023) event was held at ECMWF’s headquarters in Reading, UK, between 5 and 8 June. The event was in-person with a livestream. This allowed a record number of people to attend: around 80 came to Reading and up to 80 attended online at any one time. The event had a welcoming atmosphere and interesting discussions were had throughout. UEF events aim to provide a forum for exchanging ideas and experiences on the use of ECMWF data and products, and this was successfully achieved.
This year’s theme was ‘Ensemble Forecasting’ and focused on four thematic areas: the science of ensemble forecasting, using ensemble models and data, ensemble forecast applications and products, and communication of ensembles and probabilities. The theme was well placed as ECMWF recently celebrated 30 years of ensemble forecasting. The importance of the theme to users was clear, with many presentations talking about how ensembles are used for a large variety of topics covering meteorology, hydrology and climatology.
Meeting highlights
With Cycle 48r1 of ECMWF’s Integrated Forecasting System (IFS) being close to implementation at the time, many of the updates from ECMWF highlighted the changes and improvements brought by this new cycle. Florian Pappenberger, Director of Forecasts at ECMWF, gave an overview of the changes and showed a positive scorecard for multiple model fields. Matthieu Chevallier, Head of Evaluation at ECMWF, provided an overview of new and updated ECMWF products available in Cycle 48r1. Speaker’s Corner went into more detail, with presentations covering: meteograms, including a new visibility meteogram (Cihan Sahin, ECMWF); convective products (Ivan Tsonevsky, ECMWF); the new freezing drizzle precipitation type (Tim Hewson, ECMWF); extended-range forecasts (Fernando Prates, ECMWF); and a new snow scheme (Gabriele Arduini, ECMWF).
Stephen English, Deputy Director of Research at ECMWF, went further into the future and presented research plans for Cycle 49r1 and beyond, including 2 m temperature improvements, the possibility of hourly data assimilation updates, and major changes to the ensemble system (ENS).
Throughout the event, many areas of ensemble forecasting were presented. Chiara Marsigli (German Meteorological Service and Italy’s Arpae-SIMC) showed how ensemble resolution changes what is forecast, and how to indicate the possible occurrence of severe weather, even with low probability. Ken Mylne (Met Office, UK) gave an overview of how ensembles are exploited at the Met Office and of future plans, including supporting users with their use and communication of ensemble data. Simon Boardman (Met Office, UK) showed how ECMWF data are integrated into the Met Office system from a technical perspective and the challenges faced with this.
Communication of ensemble information to users and decision-makers is crucial. Patrick Campbell (University of Oklahoma, USA) presented the Probabilistic Hazard Information (PHI) tool and how it is used to improve the communication of tornado, hail and lightning warnings in the USA. Matteo Ponzano (Météo-France) discussed how understanding end user cost/loss ratios can improve both the products issued and user decision-making. Kosuke Ono (Japan Meteorological Agency) and Robert Neal (Met Office, UK) both presented clustering techniques as a way to summarise ensemble data for easier communication.
Alexandre Trajan (Météo-France) gave a forecaster perspective on the use of ensemble models and how they are used to provide weather warnings in France. Padraig Flattery (Met Éireann – the Irish Meteorological Service) presented a poster on the use of ensemble data in case studies of storms and heavy precipitation events in Ireland.
The prominence of machine learning has considerably increased over the last few years throughout the meteorological community. Florian Pappenberger provided an overview of what the ECMWF Machine Learning Roadmap has achieved so far and acknowledged machine learning’s busy and fast-evolving landscape. This means ECMWF's Strategy will be revised as ECMWF needs to adapt to progress in the area.
In the Ensemble and Machine Learning session, Mariana Clare (ECMWF) presented a methodology for producing reliable and skilful post-processed probabilistic forecasts without requiring ensemble information. Forest Cannon (Tomorrow.io) showed how a multi-task neural network forecast can outperform the US National Oceanic and Atmospheric Administration High-Resolution Rapid Refresh (HRRR) model. Federico Grazzini (Ludwig Maximilian University of Munich, Germany/Arpae-SIMC, Italy) presented the MaLCox model and demonstrated how it can support extreme precipitation forecasting in Italy. Linus Magnusson (ECMWF) compared the Pangu-Weather and FourCastNet machine learning models with ECMWF’s IFS. Harilaos Loukos (The Climate Data Factory, France) presented how machine learning can help to predict sub-seasonal drought and heatwaves.
An ensemble of discussions
On Wednesday afternoon, a new UEF approach was tried: 12 questions exploring forecasting, research, data, and outreach with a focus on ensembles were spread across various rooms. For each of the questions, attendees were invited to ‘vote’ with dot stickers, comment using post-its, or discuss with ECMWF experts (or a combination of the three). The aim was to generate discussion, gauge user preferences for future work, and gather general feedback. The session exceeded expectations: discussions and debate were plentiful, the sticker ‘voting’ worked very well, and a plethora of feedback was gathered.
Additionally, throughout the event, a similar approach was used to ask attendees more general questions on the ECMWF website, training, products, and the UEF, with the aim to gather feedback in these areas. Again, this provided useful ideas and feedback for future activities.
User Voice Corner
The annual User Voice Corner sent a survey to registered UEF2023 participants. The survey gathers feedback on ECMWF forecasts and forecast products. Survey responses are summarised and presented during UEF. It was clear that respondents continue to value and use medium-range ensemble forecasts, particularly for precipitation, wind and temperature, with many happy with quality and forecast skill. Some forecast issues were raised: Croatia highlighted some erroneous turbulence forecasts; Indonesia showed ‘stripy’ seasonal forecast (SEAS5) rainfall outputs; and errors in temperature were shown across the Alps, which were suspected to be due to station vs model height differences. In line with the UEF2023 theme, respondents were asked about their usage of deterministic vs ensemble data. Answers were varied: some use deterministic more, some use more ensembles, and some use deterministic at short lead times and ensembles at longer lead times. Most supported the ECMWF strategy of more ensemble-based outputs, but it was noted that this needs to be backed up with supporting activities.
Re-forecasting session
A lot of work has been done at ECMWF regarding re‑forecasting, especially given the increased resolution of medium-range ensembles and greater ensemble size of extended-range ensembles in Cycle 48r1. Magdalena Balmaseda (ECMWF) provided an overview of current configurations and options for new configurations coming in Cycle 49r1. It was shown that lagged extended-range forecasts give better skill in weeks 3 and 4, but not in week 1. The proposed configuration for SEAS6 was also shown. It provides a more comprehensive set of reforecasts as it is run more frequently, with more members and for longer than the current SEAS5 configuration. Reduced noise and improved forecast accuracy was shown, including skill to 18 months and better El Niño–Southern Oscillation (ENSO) predictability. Dominik Büeler (ETH Zurich, Switzerland) gave a user perspective showing re-forecast skill for 2 m temperature and ongoing work on defining climatological re-forecast distributions to identify extreme temperatures in individual ensemble members. The presentations were followed by a plenary session which discussed biases, flow-dependent calibration and balancing resources between real-time forecasts and re‑forecasts.
A ‘very useful’ event
Feedback during and after the event has suggested that UEF events are valuable for users to learn about the latest updates and developments in ECMWF products and services, and about future projects and areas of attention. Attendees highly valued the in-person, interactive, and networking aspects of UEF2023 as they provided opportunities for engagement across the community, strengthening cross-organisational links, and improving user understanding. The theme was described as relevant and useful, with some asking for a similar theme in the future. Feedback received included: “Thank you for organising this event, I really enjoyed it and found it very useful!” and “Thank you so much for this very inspiring meeting."