Using ECMWF's Forecasts (UEF2023)

This year's UEF theme is Ensemble forecasting. Ensembles are at the heart of ECMWF’s global forecasts at all time ranges, from the ensemble of data analysis to the medium, extended and long (seasonal) forecast ranges. ECMWF's ensemble system will undergo substantial changes in the next model cycle upgrade (48r1) this year so what better time to explore how ensembles are developed and used by the meteorology, hydrology, air quality, and climate communities.

This presentation will introduce the theme of Ensemble forecasting including providing a short history to ensembles at ECMWF.

ECMWF provides ensemble forecasts for the medium, monthly (extended-range), and seasonal range. The talk will give a quick overview of the different systems and the ensemble generation strategy. Next, we will report on the major upgrades of the medium and extended-range systems that will happen this year. The horizontal resolution of the medium-range ensemble forecasts will be increased from now 18 km to 9 km, the same resolution as the HRES forecast. For the extended-range ensemble forecast, the number of perturbed ensemble members will be increased from 50 to 100, and it will be run daily instead of twice weekly. Both changes result in significant improvements of forecast skill.

The Met Office has been investing in the development of ensemble forecasting, including use of the ECMWF ENS, and the generation of calibrated probabilistic forecasts for many years. Ensembles now use a large proportion of our investment in HPC (High Performance Computing). The science is well-proven and there is extensive evidence that probabilistic forecasts from ensembles are more skilful than deterministic models, offering higher information content to support better decisions. At the next major NWP upgrade on our new HPC we will move to an Ensemble-only NWP system. Similar to ECMWF plans, ensemble resolutions for both global and UK convective-scale forecasting will be increased to current deterministic resolutions; separate higher-resolution deterministic models will cease. Despite this progress, uptake of probabilistic forecasts in products and services is still relatively low. Recognising this, the Met Office has launched a corporate “Strategic Action” on Ensemble Exploitation to focus priorities, from model development through operational tools and into customer services, to ensure we pull through the benefits of HPC investment into better decisions which help people stay safe and thrive. We have identified several distinct use cases for ensemble forecasts to aid decision-making. Alongside the “traditional” use of probabilities, research is investigating how ensembles can be used to generate different scenarios or “story-lines”, including tools to help the operational meteorologist. There is extensive research evidence that people can make good use of forecasts incorporating uncertainty information to make better decisions, but a major obstacle remains the widely-held belief that “People don’t understand probabilities”. Another strand of the strategic action is development of training to help staff understand the opportunities offered from ensembles, how to diagnose the best information from them and how to communicate this to their customers so that the benefits can be realised.

In the recent year there has been a very fast progress from academia and the tech industry to make global medium-range weather forecasts based on Machine Learning. These models are trained on ERA5 reanalysis, and utilise a stepping procedure starting from a global conventional analysis, e.g ERA5.At ECMWF we have started to evaluate some of the ML models and compared them against ECMWF deterministic and ensemble forecasts. In this presentation we will give examples of cases such as windstorms, tropical cyclones, heat waves and cold spells, together with a first statistical evaluation for different forecast properties.

The birth of mesoscale ensembles with limited-area models, aiming at a quantification of the forecast uncertainty for severe weather, followed closely the development of global ensemble forecasting. Nowadays, many centres operationally run ensembles at km-scale, which resolve explicitly the convection. The higher resolution and the stochastic behaviour of the convection introduce new challenges for an optimal usage of these forecasts. The forecast uncertainty is combined with the spatial uncertainty typical of high-resolution predictions, therefore different post-processing, products and verification methods are needed. In this contribution, the specifics of high-resolution ensemble predictions will be presented, analysing examples of their usage and the related challenges. In particular, it will be highlighted how ensembles are utilized for predictions of high-impact weather, discussing their strengths and shortcomings.

The Copernicus Climate Change Service (C3S), implemented by ECMWF on behalf of the European Commission since 2015, has included a seasonal forecast component since its inception. It consists in a multi-system framework, currently including eight ensemble forecast systems from operational centres around the world. C3S collects, processes and offers users data from these predictions - as well as corresponding past predictions, to allow evaluation and calibration of real time forecast products - and produces and displays a set of graphical products for a subset of the available variables.
This set up will be described, with highlights on recent products and plans for the near future.

I created a poster to show the usage of ECMWF's ensemble model data at DWD (German Weather Service), please see attachment

The multi-model ensemble system for impact-based forecast was developed by using the following 7 numerical prediction models.
(1) ECMWF global model (9km 137 levels)
(2) ECMWF global ensemble model (18km 137 levels 51 members)
(3) KIM (Korean Integrated Model) global model (12km 91 levels)
(4) UM (Unified model) global model (10km 70 levels)
(5) UM global ensemble model (32km 70 levels 25 members)
(6) UM local model (1.5km 70 levels)
(7) UM local ensemble model (2.2km 70 levels 13 members
The impact-based forecast guidance was produced from 93 members of the
multi-model ensemble by using Generalized Extreme Value (GEV) probability distribution curve. Also bias correction for 10 days which uses ‘decaying average method’ was applied. The heat wave forecast guidance is divided into 4 categories: ‘concern’,‘caution’,‘warning’, and ‘alarm’ according to the probability value for each temperature.
The performance of the impact-based forecast guidance developed from multimodel ensemble was evaluated. According to verification results, it was found that the performance for warning and alarm level, that are high risk levels was improved compared with impact-based forecast from the deterministic forecast.

Since agriculture is highly exposed to weather-related risks such as drought constraints, Decision Support Tools (DSTs) are now frequently used in irrigation management. The current use of DSTs mainly relies on deterministic weather forecasts that do not account for the associated weather uncertainties. Few irrigation DST users take uncertainty into consideration by utilizing ensemble of historical weather observations (EHO) as input to their models.
In this work, we introduce a novel approach that uses IFS-EPS as input to a vine irrigation DST, named WaLIS, that takes agro-meteorological data (T, RR24, ETP) as input and computes a crop water stress index as output. Our results show that this approach has significant better probabilistic performance than the use of EHO as input to the WaLIS model.
We also investigate the effect of statistical post-processing on the probabilistic skill of the water stress index with two different approaches. In the first case, we directly post-process the output of the WaLIS model. In the second case, only agro-meteorological forecasts are post-processed, before being used as inputs of the WaLIS model. Our results show that locally in 4 out of 10 tested sites in southern of France, the raw ensemble forecasts could be significantly improved by the post-processing, and that generally there is no significant performance difference between the two competing post-processing approaches.

The Met Office’s probabilistic weather pattern forecasting tool, Decider, is a medium-to extended-range forecasting tool based on a set of 30 predefined weather patterns for the UK and surrounding European area. It is designed to help summarise key aspects from the large volumes of data which ensembles provide, such as through helping meteorologists communicate the most likely weather pattern transitions, assess forecast confidence and assess forecast consistency. Once weather pattern characteristics are understood, in terms of their climatologies or impacts, it then becomes possible to interpret forecast output and describe likely consequences. Daily weather pattern probabilities are based on the number of ensemble members objectively assigned to each weather pattern definition, with output available for individual models (MOGREPS-G, GEFS, ECMWF medium-range, ECMWF extended-range and GloSea6) as well as using a new seamless blended multi-model configuration. Objective verification shows that the new multi-model forecasts are at least as skilful as the best performing individual model. Results also show that varying model weights within the blend has little impact on forecast skill. Having a single seamless forecast not only has the benefit of improved forecast skill for many weather patterns, but it also helps speed up the decision making process for operational meteorologists. Examples of Decider’s new multi-model use operationally highlight the different ways the probabilistic forecast visualisations have aided communication of the key weather story. Future changes to the multi-model blend will pull through benefits of the new daily updating 101-member extended-range ensemble from ECMWF.

The correct prediction of intense precipitation events is one of the main objectives of operational weather services. This task is even more relevant nowadays, with the rapid progression of global warming which makes these events more intense. Numerical weather prediction models have improved continuously over time providing uncertainty estimation with dynamical ensembles. However, direct precipitation forecasts are still challenging. Machine learning in this context could open a new approach of a non-linear user oriented post-processing. Here we describe a specific post-processing chain, based on a random forest pipeline, specialised in recognizing favorable synoptic conditions leading to precipitation extremes and subsequently classifying them in predefined types. The application focuses on northern and central Italy, taken as a testbed, but extensible to others regions. The system, which is called MaLCoX (Machine Learning model predicting Conditions for eXtreme precipitation) and is running daily at ARPAE, has been trained with the ARCIS gridded high-resolution precipitation dataset as the target truth, and with the ECMWF reforecast dataset over the last 20 years as the input predictors. We show that, with a long enough training period, the optimal blend of large scale information with direct model output improves the probabilistic forecast accuracy of extremes in the medium-range. In addition, with specific machine learning methods, we provides a useful diagnostic to explain to the forecasters the underlying physical storyline which make a particular meteorological event develop into an extreme. This additional diagnostic complement and augments even further the value of standard EPS probabilities and Extreme Forecast Index.

ECMWF ensembles are under-dispersive, and whilst c48r1 is reported to increase the ensemble spread, it does not look to solve the issue.
In some situations (e.g. daytime high temperatures) there is a known issue with the model, and usually improvements are in progress.
In other situations (e.g. the US 'Christmas storm') there seems to be a lead time beyond which the forecasts (op and/or ensembles) fail to forecast an event.
Is this because of
a) model error (there is no initial condition at lead time x which would produce a reasonable forecast) or
b) ensemble perturbation choice (there are initial conditions at lead time x which would produce a reasonable forecast, but they weren't chosen) or
c) some combination?
In this talk we'll look at a couple of situations where the forecast suddenly improved as the event got closer, and try to ascertain what was limiting the skill at longer lead times.

[Another reason why a move from bi-weekly to daily extended forecasts is a good thing for an operational forecaster!]

We'll then discuss at how understanding the cause of such failures can help in operational forecasting and decision making, and future NWP system (model + ensemble choice) development.

In October 2018, Ireland’s meteorological service Met Éireann implemented the Irish Regional Ensemble Prediction System (IREPS). Met Éireann's ensemble modelling system is designed to improve the accuracy and reliability of weather forecasts, particularly in situations where uncertainty is high, by providing probabilistic forecasts, particularly for severe weather. The use of ensembles also improves our impact-based forecasting capabilities, improving the accuracy of our weather warnings.

IREPS uses the HARMONIE-AROME model configuration of the shared ALADIN-HIRLAM NWP system. It consists of 15 ensemble members with 1 control member. The control member produces 54-hour forecasts and runs eight times daily at 3-hourly intervals from 00Z to 21Z. Each of the 15 ensemble members produce 57-hour forecasts, with members 1-10 running at 6-hourly internals from 00Z and members 11-15 running at 6-hourly intervals from 03Z. A combination of members is used to create a semi-continuous lagged EPS at 3-hourly intervals. IREPS runs with a horizontal grid spacing of 2.5 km, a spatial extent of 1000 x 900 grid points, and 65 vertical levels. Both conventional (observations from weather stations, ships, buoys, balloons and aircraft) and scatterometer (satellite wind data) observations are assimilated.

IREPS ensemble members are constructed by perturbing the boundary conditions using the Scaled Lagged Average Forecasting (SLAF) technique. Also, surface perturbations are applied to several variables within the surface code in the model such as sea surface temperature and soil moisture. These perturbations are intended to account for uncertainty related to energy transfer from the model surface to the planetary boundary layer (usually the lowest kilometre of the atmosphere).

In 2022 the IREPS system was launched to the public on Met Éireann’s website, where we display the percentage probability of forecast precipitation for 6 and 24 hours, mean hourly 10-metre wind speed and gust, and 2-metre temperature below or above certain thresholds, for up to 2 days ahead.

Met Éireann also use ECMWF-EPS for forecasts beyond 2 days ahead, for warning probabilities, meteograms and mean rainfall maps. One example of the use of ECMWF-EPS is a forecast of average, median, 85% and 25% percentile cloud cover percentage, precipitation, windspeed and temperature for racecourses around Ireland. Another example of the use of EPS is providing national emergency agencies with the probabilities of exceeding our major warning thresholds for the week ahead.

Over the coming years, Met Éireann aims to enhance our use of ensemble forecasts to improve our dispersion modelling capacity by using output from both IREPS and ECMWF-EPS to improve modelling of the dispersion of nuclear material, vectors and pathogens which can carry animal and plant diseases (e.g. biting midges, beetles), volcanic ash, Saharan dust, pollen and fungal spores.

By harnessing the power of ensemble forecasting, Met Éireann is working to ensure that stakeholders and the public have the information they need to make informed decisions about their daily activities and long-term planning.

The persistence of a given atmospheric state can provide information of crucial importance for forecasting extreme weather events. Persistence, however, is often described in somewhat subjective terms. Dynamical systems theory has been used to quantify atmospheric persistence in an objective manner, and this metric has been shown to be associated with temperature extremes in Europe. We apply this metric, to the best of our knowledge, for the first time to reforecast data. Specifically, we assess the applicability of a dynamical systems persistence metric to extended-range forecasts, discussing potential links to heatwaves in Europe and its potential value as a forecast evaluation metric.

MeteoSwiss uses ensembles of ECMWF and of local area models routinely in providing weather forecasts to the general public and as basic information for weather warnings. Still, uncertainty in forecasts is often not included in standard forecast products due to the communication challenge of conveying the more complex information from ensemble forecasts. We are convinced that clever visualizations can make a significant contribution here and have developed some examples as an initial exploration into the possibilities. These representations strive to be intuitively understandable while still providing users with useful insights regarding forecast reliability. We believe UEF is the perfect place to invite feedback from both ECMWF users and staff and we welcome ideas or experiences related to probabilistic forecasting visualization concepts. Our next step is to select a few promising approaches before collaborating with a professional designer to further refine them.

Seasonal meteorological forecasts can provide valuable information to manage extreme events, such as activating drought early warning measures or informing reservoir operations. Since these forecasts might have substantial biases for certain areas, post-processing of raw predictions can be crucial to provide adequate meteorological information. One of the goals of the WATER4CAST (Integrated Forecasting System for Water and the Environment) project is to explore adequate post-processing methods to improve the quality of seasonal meteorological forecasts. To achieve this objective, an innovative post-processing method based on fuzzy logic has been applied and compare with alternative procedures applied to the Jucar River Basin District (JRBD).

Seasonal forecasts from the Copernicus Climate Change Service (C3S) have been used to test the potential of fuzzy logic, in particular forecasts products from ECMWF (SEAS 5), Météo-France (System 8), DWD (GCFS 2.1), and CMCC (SPS v35) for the 1995-2014 period. The variables analysed include precipitation, average, minimum and maximum temperatures, solar radiation and wind speed. ERA5 time series were used as reference for post-processing. The performance of fuzzy logic post-processing has been compared with linear scaling and quantile mapping employing the MAESS and CRPSS skill indices. Results show how the forecasting skill changes after post-processing for each method, the resulting forecast skills per variable, forecast months and lead month, and how fuzzy logic performs compared to linear scaling and quantile mapping.

The Met Office has deployed the Standard Gridding Engine to translate ECMWF model output to be compatible with Met Office Post-Processing systems. Primarily ECMWF data will be incorporated into the IMPROVER probabilistic post-processing system, extending its forecast from 7 to 14 days.

The Standard Gridding Engine, routinely called StaGE, is a Python library developed within the Met Office. It is designed to ‘standardise’ raw NWP model output to provide both a decoupling – insulating users from specifics of the model – and improve consistency between the data offered from different models. The process covers file format, metadata, set of diagnostics, horizontal grid and map projection, vertical level types and sets, and time steps .

StaGE is built around the Scitools Iris software package, with a translation layer for utilising ECMWF grib data provided by the iris_grib and ecCodes packages. Users define Diagnostics in configuration files using a series of plugins that perform the standardisaton steps. For ensemble data this process is carried out separately for each member and the resulting data combined so that users can access a single "cube" of standardised data for a given diagnostic at a given forecast time point by loading one NetCDF file.

StaGE has been successfully deployed within an ecFlow suite on the ECMWF HPC. It is processing ECMWF ensemble data from the Integrated Forecasting System (IFS) twice a day for all 51 members. The results show that the system is able to standardise IFS output to match that of the Met Office Unified Model.

The StaGE output is currently being used by Met Office scientists to incorporate ECMWF data in post-processing systems in combination with data from the Unified Model. The system has the potential to improve the accuracy of Met Office forecasts by providing access to high-quality ECMWF data for probabilistic post-processing systems and enabling an extended forecast out to 14 days.

Seasonal forecast is an early warning system that contributes to anticipatory management by providing spatial and temporal information of the near future. This study examined the skill of ECMWF system 5 (SEAS5) sub–seasonal–to–seasonal (S2S) forecasts over Mainland Southeast Asia (MSEA). We evaluated the SEAS5 skill of temperature and precipitation for 30 years (1985–2014) against two reference model datasets, WFDE5 and APHRODITE, using probabilistic forecast verification skill metrics at grid cells for each month. Then, the SEAS5 data was used to force the Variable Infiltration Capacity (VIC) hydrological model to predict runoff and streamflow. These hydrological results were compared against the WFDE5-driven streamflow reanalysis and observed station data, using the same probabilistic skill statistics. The results show a prediction potential for temperature beyond 2 months in advance. The skill of precipitation and streamflow forecasting is limited to the first month. Strong seasonal and regional dependence occurs. The model shows high forecast skills during the pre-monsoon (April–May) and post-monsoon (October–November), arguably the period when its usefulness is potentially highest. Conversely, poor skill is observed during the rainy monsoon season (June–August). In eastern and southern MSEA, i.e. in eastern Thailand, Cambodia, Vietnam and Malaysia considerable skill levels are found. Year–to–year precipitation tercile plots highlight skill in predicting the anomalous seasonal conditions caused by the ENSO. Overall, SEAS5 and derived hydrological forecasts show useful skill that can potentially be used for hydrological and agricultural anticipatory management in this region.

When modelling possible future renewable electricity systems, a strong focus needs to be directed to the input weather variables driving any such system. Since we cannot know the exact weather in any slightly distant future, a probabilistic approach is usually chosen, modelling the system over many possible scenarios, typically all of the past recorded weather data available. However, this narrows the range of situations considered to about 40 years, placing fundamental limits on the analysis, e.g. of rare, extreme scenarios.

In my work, I explore the possibility of using past expired ensemble forecasts from the ECMWF ENS Extended [1] to drastically increase the number of scenarios considered to up to 10 000 years of data. These ensemble forecasts are physical models that are regularly initialized from the same slightly perturbed snapshot, but due to the chaotic nature of weather, their predictions diverge from each other. The later stages of their predictions are thus entirely independent predictions of what the weather could have been, including the correct spatial correlations. I analyze the data from the operational archive of ECMWF to assess their suitability for modelling renewable systems of the future and demonstrate how this wealth of additional weather scenarios can enable the utilization of otherwise heavily data-dependent machine learning techniques in energy modelling.

[1] European Centre for Medium-Range Weather Forecasts (ECMWF) Atmospheric Model Ensemble extended forecast https://www.ecmwf.int/en/forecasts/datasets/set-vi

Cluster analysis can aggregate many forecasts contained in an ensemble forecast into a small number of forecast scenarios. This is beneficial for time-sensitive operational forecasters in that they can easily grasp multiple possible forecast scenarios. In addition, since the current short-range operational forecasts at the Japan meteorological agency (JMA) is based on deterministic forecasts, the provision of a small number of scenarios through cluster analysis is considered to have greater affinity and availability than probabilistic products. When considering cluster analysis results in the time direction as a possible weather scenario, it is necessary to consider changes in the cohesiveness of ensemble members within a cluster. In this study, we divided 4 scenarios from the 500-hPa geopotential height forecasts by clustering every 3 hours up to 39 hours ahead using the mesoscale ensemble prediction system (MEPS) at the JMA with the 5 km grid spacing and the ensemble size of 21. Also, we linked the clusters in the time direction by focusing on the membership that constitutes a cluster at adjacent forecast lead times. This method provides the well-divided 4 multiple scenarios from MEPS at each lead time and the cohesiveness information of the members in the cluster. The presentation will provide an overview of this method and introduce the characteristics of multiple scenarios obtained from this method.

The drought of summer 2022 is considered as one of the most severe episodes on the whole French territory, as 2003 or 1976. In order to better support agricultural stakeholders, the agricultural technical institutes (applied research center) and the Météo-France research centre proposed a mapping of various very simple probabilistic indicators based on the probabilistic precipitation forecasts from the IFS-EPS (Integrated Forecasting System Ensemble Prediction System) model. The aim was simply to provide information on the probability of reaching a certain rainfall accumulation (10mm, 15mm and 30mm) over the next 7 days and the following week as well.

This information on rainfall forecasts and associated uncertainties was disseminated in the form of a bulletin containing maps and an interpretation, at a rate of 3 editions per week. These bulletins were published on the web (http://www.modelia.org/moodle/course/view.php?id=86), and social networks (LinkedIn and Twitter) were mobilised to promote wide distribution. A mailing list was set up to quickly gather more than 300 recipients, the majority of whom were farmers (42%) or agricultural advisors (22%).

Crop and livestock experts also used this information to propose technical advice to help farmers adapt to this extreme event. This was the case in particular (i) for decision support for rapeseed sowing in August by coupling data with agronomic decision rules (need for rainfall accumulation according to the structure of water status of the seedbed) and (ii) for the management of sowing of grasslands.

Thus, this large-scale experiment showed the contribution of a probabilistic approach to rainfall accumulation over a 2-week horizon for farmers and agricultural advisors in order to help them make decisions, particularly during such a major climatic crisis as this drought episode.

It is planned to extend the experimentation to 2023 before further evaluating the added value of the system in terms of prediction quality and the formatting of probabilistic information.

Most of our users and customers are not familiar with statistics and probabilities. They often need a deterministic weather scenario to make their decisions: trigger a procedure, cancel an event, mobilise personnel, etc.

When the forecast is uncertain, ensemble models allow the forecaster to quantify the uncertainty, define ranges of values, talk about the possibility of an extreme event or an alternative scenario.

By using ensembles, the forecaster moves from uncertainty of a forecast to forecasting uncertainty. This, combined with a good understanding of the user's problems, allows the right decision to be made.

Here are some examples of the methods used at Météo-France.

The Japan meteorological agency has operated a regional mesoscale model (MSM) with a horizontal grid spacing of 5 km to forecast severe weather phenomena. Also, a mesoscale ensemble prediction system (MEPS) based on the MSM has been operated to quantify the uncertainty in initial and lateral boundary conditions of the MSM with the ensemble size of 21. We have performed a cluster analysis on the 500-hPa geopotential height (Z5) forecasts of these 21 members and created four representative forecast scenarios that take into account the temporal variation in membership within a cluster (see the author's oral presentation). In this study, we have tried to select the cluster with Z5 forecast errors smaller than the MSM at 36 hours, focusing on ensemble members with small Z5 forecast errors at 6 hours after the initial time. As the clustering method in this study provides temporal variation of intra-cluster membership, it also provides important practical information on how long the members with fewer errors after 6 hours will constitute the same cluster in the future. The presentation will demonstrate the usefulness of intra-cluster membership through a case study in which the optimal scenario was successfully selected, and will discuss the reasons why the optimal scenario was able to be selected from the perspective of linearity of error growth. The characteristics of precipitation forecast scenarios constructed from the four clusters of the Z5 will also be presented.

Extreme precipitation events during spring and early summer can induce hazards of different forms and associated impacts over the countries in Central Asia. Therefore, reliable forecasting of heavy precipitation is essential to generate accurate warnings for debris flows in the study area. This study will focus on two weather forecast products developed by the European Centre for Medium-Range Weather Forecasts (ECMWF), a standard Ensemble Forecasts (ENS) and post-processed ecPoint Rainfall, to predict the risk of debris flows in the region. Furthermore, important questions of forecast performance and debris flows occurrence dependency, for example, forecast verifications and precipitation-induced debris flow thresholds in Central Asia, are investigated. In this way, the study discusses designing and maintaining effective early warning systems for debris flows in Central Asia and the implementation of impact-based forecasting in the region. The results from ENS and ecPoint Rainfall forecasts will provide new resources and a more detailed investigation of precipitation-induced hazards in the study area.

Some operational ensemble forecast products developed at Météo-France and addressed to commercial, industrial and institutional end-users will be described during the presentation. Most end-users are keen to exploit ensemble forecast products for decision-making, such as workflow planning, human resources management, preventive actions. An improper use of a probabilistic product can lead to taking bad decisions, which may result in consistent economic loss. Stakeholders using ensemble forecast product may be classified into two categories: the ones who integrate probabilistic products into decision-making models or leverage human expertise to provide valuable information, the ones who are not able to exploit probabilistic information and need a translation to binary forecasts. One of the current objectives of Météo-France is to enhance the current range of products and services offered by the public institution by integrating uncertainty, reliability and probabilistic forecast information in order to meet the specific needs of each end-user. A new approach has been explored in order to put the focus on the need of the end-user. The production and the commercial departments launched a testbed in collaboration with wine companies to provide a seamless ensemble forecast product to protect vineyards during spring frost. The aim is to process ensemble forecasts and customised economic cost/loss functions to produce relevant and easy-to-use risk forecasts. Future plans include extending the employment of this approach to other commercial and institutional sectors.

Convective storms can lead to several forms of severe weather, such as tornadoes, flash floods, damaging wind gusts, and/or large hail. Thus, it is essential to assess the performance of predictors commonly used by meteorologists in forecasting such storms. The presence of conditional or convective instability, sufficient moisture, and an uplift source is necessary for the onset of convective storms. In addition, the severity of these storms depends on the relationship between these ingredients and vertical wind shear. This study evaluates the performance of several predictors based on the ECMWF deterministic and ensemble forecasts. These predictors include Convective Available Potential Energy (CAPE), convective inhibition (CIN), vertical wind shear, cloud base height, among others. The performance of these predictors is illustrated for four events occurring at distinct seasons.

The first two events focus on hailstorms that caused extensive losses to the agricultural sector in northern and central regions of mainland Portugal, damaging several thousand hectares of vineyards and olive groves in the first event and 600 hectares of cherry and peach trees in the other. In the latter event, 68 goats were also killed by lightning.

The two other episodes refer to tornadic storms that caused damage to vehicles, vegetation and various facilities in mainland Portugal. In the first episode, an F1 tornado hit the north coast while another affected the south coast. The former was produced by a quasi-linear convective system (QLCS), while the latter was produced by a low-altitude mini-supercell. In the last episode, a supercell embedded in an instability line caused flash flooding in the Lisbon area. This supercell also spawned two F1 tornadoes, and the one with the shorter path (1.5 to 2km) hit Lisbon.

We present a machine learning model that post-processes high-resolution, deterministic forecasts to produce short to medium-range probabilistic forecasts for seven core weather variables. We developed and operationally implemented a multi-task neural network with a custom loss function, namely the Continuous Ranked Probability Score (CRPS). We tested this methodology using raw deterministic ECMWF HRES forecasts, ERA5 reanalysis fields as target data, and ERA5 invariant fields as supplementary features. Additionally, we tested and operationalized this model using data from the High Resolution Rapid Refresh (HRRR) model as input. This technique combines the strengths of high-resolution Numerical Weather Prediction (NWP) modeling with complex non-linear machine learning to generate more accurate deterministic forecasts alongside probabilistic forecasts, adding substantial predictive and actionable information. The results show deterministic forecast improvements in Root Mean Square Error (RMSE) over the HRRR from 1% to 12.5%. Using the CRPS as a metric to validate the probabilistic forecasts, we find a 22% to 38% improvement over the HRRR model. In addition to this increased forecast skill, the multi-task neural network approach is affordable to train and lightweight enough to run operationally on hourly forecasts. While our application used 21 ensemble members, this machine learning based approach has the flexibility to generate any number of ensemble members to best fit distribution of the forecast variables without significantly increasing computational costs.

Since 2014, Indonesia Agency for Meteorology, Climatology, and Geophysics (BMKG) has been using the output of seasonal predictions of the European Centre for Medium-Range Weather Forecasts (ECMWF) as based of seasonal rainfall prediction (up to the next three months). In this study, we assess the skill of SEAS5. Prediction skills are assessed using several indicators, namely: correlation, relative bias (%), and the model's ability to simulate the wettest and driest months. We used ECMWF reforecast data (ensemble of 51 members) for the period 1981-2016 and observation data from 96 stations over Indonesia. In general, we found that SEAS5 well replicates the annual cycle pattern at 82 stations, with a correlation of more than 0.75. However, SEAS5 was only able to accurately predict the occurrence of the driest and wettest month of the year at 45 stations.

SEAS5 shows a higher correlation in the peak dry season (August to October), compared to the peak rainy season (December to February), especially for the southern part of Indonesia, with the highest (lowest) correlation of 0.6 (0.1) in October (January). For the eastern part, where the annual rainfall pattern is opposite to the southern part, SEA5 has a higher correlation in the peak of the rainy season (July to September) than the peak of the dry season (December to January). Meanwhile, in the western part, where the annual cycle has two rainy peaks, the skill is relatively low; the highest average correlation is around 0.38 in October. Calculation of relative bias for the whole Indonesia region shows that the largest bias generally occurs in November to April with an average of 34%, and the lowest bias occurs in May to October with a bias of around 4%. We found that SEAS5 tends to have good skills for southern and eastern Indonesia.

Sub-seasonal and seasonal forecasts provide opportunities for strategic advice for the agricultural sector. In Bangladesh, a team consisting of Weather Impact, Wageningen Environmental Research and Digital Innovation for Impact have implemented an operational S2S forecasting system at the Bangladesh Meteorological Department (BMD).

Forecasting at S2S timescales needs to be done using ensemble modelling. To identify which models to use in Bangladesh, an extensive skill assessment is done using a wide range of models. The ECMWF model is the best performing single model, but including other models into a multi-model ensemble increases the forecast skill. The final combination is calibrated using a gridded observational dataset from the BMD. Since March 2023, the S2S forecast system is operational in experimental mode and forecasts are issued on a weekly basis.

The S2S forecasts that are issued by the BMD are used by the Bangladesh Department of Agricultural Extension (DAE). The DAE is translating the meteorological forecasts into agricultural advice, which is disseminated to farmers using the extension network. The challenge of the translation is to get valuable and actionable advisories from the forecasts with a lower skill than short- to medium range forecasts. The nature of the advice will therefore be more in preparation and strategic domain.

The final challenge in this forecast implementation is the sustainability of the complete system. The procedures needs to be institutionalized at the BMD and DAE. Agreements on roles and responsibilities are made under the supervision of the international implementation team.

In the presentation, Bob Ammerlaan, project coordinator of the project, will provide insights in the journey we took to development this type of forecast system and locally implement it in Bangladesh. He will also share multiple use cases of how S2S ensemble information can be used in agriculture.

The main objective of the H2020 Climate Intelligence (CLINT) project is the development of an Artificial Intelligence framework composed of Machine Learning techniques and algorithms to process big climate datasets for improving Climate Science in the detection, causation, and attribution of Extreme Events (EEs), namely tropical cyclones, heatwaves and warm nights, droughts, and floods. The CLINT AI framework will also cover the quantification of the EEs impacts on a variety of socio-economic sectors under historical, forecasted, and projected climate conditions, and across different spatial scales (from European to local), ultimately developing innovative and sectorial AI-enhanced Climate Services. Finally, these services will be operationalized into Web Processing Services, according to the most advanced open data and software standards by Climate Services Information Systems, and into a Demonstrator to facilitate the uptake of project results by public and private entities for research and Climate Services development.

The presentation gives a brief overview of the CLINT project (15 partners among which 2 operational centres) and then focuses on presenting the challenges faced in implementing a cloud based operational demonstrator of AI-enhanced sub-seasonal and multi-system seasonal forecast aiming at forecasting heat waves and warm nights and their attribution to climate change. Three categories of challenges are discussed: scientific (methods for improving the detection, forecast, and attribution of the event), technical (data availability, data handling and processing, and overall coherency) and computational (performance/cost ratio and cost optimisation on a public cloud).

Low-visibility conditions (LVC) are a common cause of air traffic, road, and sailing fatalities. Forecasting those conditions is an arduous challenge for weather forecasters all over the world. In this work, a new decision support system is developed based on an analog ensemble (AnEn) method to predict LVC over 15 airports of Morocco for 24 forecast hours. Hourly forecasts from the AROME model of eight predictors were used to select the skillful analogs from 2016 to 2018. The verified hourly observations were used as members of the ensemble. The developed ensemble prediction system (EPS) was assessed over 1 year (2019) as a single-value forecast and as a probabilistic forecast. Results analysis shows that AnEn outperforms persistence and its best performances are perceived generally during night and early-morning lead times. From continuous verification analysis, AnEn forecasting errors are found to be location- and lead-time-dependent and become higher for low-visibility cases. AnEn draws an averaged Centered Root Mean Square Error of about 1500 m for all visibilities, 2000 m for fog and 1500 m for mist. As an EPS, AnEn is under-dispersive for all lead times and draws a positive bias for fog and mist events. For probabilistic verification analysis, AnEn visibility forecasts are converted to binary occurrences depending on a set of thresholds from 200 m to 6000 m by a step of 200 m. It is found that the averaged Heidke Skill Score for AnEn is 0.65 for all thresholds. However, AnEn performance generally becomes weaker for fog or mist events prediction.

Prior to UEF 2023 all those who had registered to attend were invited to complete an online survey to provide feedback to ECMWF on its forecasts and its products. Some of the questions were standard questions that we ask every year. Other questions were more focussed on this year's "Ensemble Forecasting" theme. This wide-ranging presentation will summarise the results of the survey, and will include scientific and technical responses from ECMWF regarding some of the points raised.

This presentation will give an update on ECMWF research. it will introduce science improvements in new versions of the IFS code system planned in 2023-4 and describe some likely avenues looking further into the future. The focus will be on elements most likely to impact users.

Cycle 48r1 is the first science upgrade on the ATOS machine in Bologna and will include a major improvement in the ensemble forecast capability. Firstly, the resolution for the medium range forecast will improve to TCo1279, equal to the current HRES resolution. Secondly, the extended range forecasts will run every day and with 101 members (up from 51). In addition to these headline grabbing changes, 48r1 is a large and complex cycle, with many other science changes. Notably the multi-layer snow model, more use of satellite observations over land and improved ozone modelling.

Cycle 49r1 also has a strong focus on improving ensemble forecasts. The move from the SPPT to SPP perturbations, follow many years of careful research, introduces a more physically based approached. The ensemble of data assimilations (EDA) will move to "soft-recentring" which allows a more accurate solution at lower cost, opening the way to improving the EDA affordably. 49r1 also will bring a number of changes that will improve 2m temperature forecasts (on top of benefits in this area from the resolution change in 49r1): in particular Synop assimilation and new land use maps. Cycle 49 is also envisaged as the cycle for the next reanalysis, ERA6, and SEAS6 systems, although it is likely this will require an extra interim cycle, 49r2, to be built to enable use of NEMO4 and the SI3 models, which may not be ready in time for 49r1. There are also many exciting new observations coming online in this period. Other contributions in 49r1 will be also briefly mentioned.

The talk will end by looking further ahead, to research that will deliver to operations in 2025 or later. Key elements include more use of Machine Learning, unified land data assimilation, higher resolution data assimilation, more and improved atmospheric composition, GPU adaptation and making full use of next generation observations such as second generation European polar satellites (EPS-SG) and Third Generation European geo satellites (MTG).

An overview is given about recent developments in the performance of ECMWF’s operational forecasting system. Time scales from medium through extended range to seasonal are considered and forecast skill improvements due to recent model upgrades are highlighted. Comparisons with other centres are made, and ECMWF’s lead in upper-air and surface forecasts is analysed. Following the theme of this UEF, the focus of the talk is on ensemble verification results. Forecast reliability and resolution are evaluated, and the effect of accounting for representativeness in the ensemble verification against observations is shown. New results on the use of gridded observational datasets for precipitation and radiation fluxes are provided as well. A brief look at trends in extended-range forecast skill concludes the presentation.

Early warning systems (EWS) play a crucial role in the preparation for natural hazards and mitigation of their devastating impacts, and with flooding the cause of nearly 40% of all natural hazards since 2000 (Guha-Sapir et al., 2018), efforts have been made to better capture uncertainty associated with the forecasting of complex hydrological processes. The European and Global Flood Awareness Systems (EFAS and GloFAS) from the Copernicus Emergency Management Service have been pioneers in the use of ensemble forecasts, moving away from traditional deterministic approaches to provide more beneficial probabilistic flood forecasts (Wu et al., 2020). They leverage ensemble numerical weather prediction (NWP) meteorological forecasts (ECMWF-ENS, and in the case of EFAS also COSMO-LEPS) as input to a fully calibrated hydrological model (LISFLOOD) to produce an ensemble of daily hydrological forecasts for lead time up to 30 days (GloFAS), updated twice daily in the case of EFAS. The poster illustrates the use of ENS forcing datasets in the EFAS and GloFAS operational flood forecasting systems and shows how the hydrological forecast products can effectively communicate uncertainty in the forecasts for selected flood events.

Flash floods are one of the most devastating natural hazards. They can occur in very large or small rural or urban areas, with little to no warning. Extreme (localized) rainfall plays a crucial role. This presentation compares the rainfall forecast performance, for the raw ECMWF ensemble (ENS) and post-processed point-scale output derived from that (ecPoint), in pinpointing areas at risk of flash floods. Performance evaluation is based on location and timing accuracy for the flash floods. Long-term objective verification and case studies are used to compare. Although ENS effectively identifies areas at flash flood risk in instances of large-scale rainfall, its performance falters when confronted with localized extreme convective systems. We show how ecPoint yields superior results for both scenarios, pinpointing well areas at flash flood risk up to medium-range timescales. This enables decision-makers to extend their preparedness and action time window. This presentation will also demonstrate forecast system strengths and weaknesses and how forecasters can leverage these to produce better predictions of areas at flash flood risk up to medium-range lead times.

Machine learning techniques are increasingly used in weather forecasting, to either improve results from NWP models or emulate some aspects of the forward integration part using a purely data-driven approach. Whilst methods that attempt to fully emulate NWP models are both very complex and computationally expensive, much simpler and computationally cheaper techniques can be used to create a probabilistic forecast from a single global high-resolution forecast by post-processing. Specifically, we focus on the relatively novel technique of a Bayesian Neural Network and show how it can predict the distribution of the forecast error relative to its own analysis. By adding these error distributions to the original forecast, we can create a probabilistic forecast.

This methodology is particularly useful for very high-resolution NWP models where running an ensemble is too computationally expensive and for machine learning approaches where no uncertainty information is available. We show how this methodology can be successfully applied to both ECMWF’s high-resolution forecast and a purely data-driven weather forecast model being run at ECMWF, for both the surface variable of 2m temperature and the atmospheric variable of Z500. These probabilistic forecasts have been verified using standard metrics and show good reliability and improved skill relative to the original forecast, at the lead times tested.

In this presentation, I will go through recent and ongoing forecast product development activities at ECMWF, in response to user requests and feedback and as part of our contributions to EU programs Copernicus and DestinE.
With IFS Cycle 48r1, the horizontal resolution of the medium-range ensemble will increase from 18 to 9 km - the same horizontal resolution as the HRES. Cycle 48r1 will also bring a major upgrade to the configuration of the extended-range ensemble, which will be a completely separate system running daily from 00 UTC out to day 46 with 101 members. Upgrades and enhancements to existing products will be introduced, including more physically consistent parameters for convection, the introduction of a new precipitation type, new meteograms and a set of extended-range forecasts generated daily.
I will also review other developments underway, especially in the area of environmental forecasting, high-impact weather forecasting, and in preparation of forthcoming IFS cycles for NWP, environmental and climate predictions.

Current warnings issued by the US National Weather Service advise the public regarding whether or not they are at risk from a severe weather event. This system of deterministic warnings undoubtedly protects lives and property, however it is also recognized that it allows for significant inequalities in both the usefulness and timeliness of critically important information provided to those in need. Probabilistic Hazard Information (PHI), which is an important part of NOAA's Forecasting a Continuum of Environmental Threats (FACETs) paradigm, expands upon the current communication of weather hazards by providing likelihood information with additional spatial and temporal precision relative to traditional weather warnings. When provided by forecasters, this additional information potentially allows decision makers and the general public to gain a much more comprehensive understanding of weather threats and, as such, better address a larger range of societal needs. In an effort to advance PHI-related research, frequent experiments carried out at the Hazardous Weather Testbed (HWT) at the National Severe Storms Laboratory (NSSL) have provided feedback, direction, and inspiration for several PHI-related concepts.

This presentation will provide a brief history of PHI research, cover the concepts tested in recent HWT experiments, and will discuss how the findings from those experiments have shaped ongoing and future research. Topics covered will include the automated computer models and guidance that assist with PHI creation, the software developed for PHI generation, the Threats-In-Motion (TIM ) guided warning paradigm, and different methods of communicating PHI. Importantly, findings regarding the feasibility of the continuous generation of severe weather PHI and of successfully communicating PHI to decision makers and the public will be discussed. A summary of the steps currently being taken to operationalize PHI and TIM will also be presented.