Cristina Primo
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Cristina Primo has a total of 72 co-authors. Here are the top co-authors by number of shared publications:
- Manuel Baumgartner
Institute for Physics of the Atmosphere, Johannes Gutenberg University, Mainz, Germany
6 shared publications - Sebastian Lerch
Karlsruhe Institute of Technology
5 shared publications Aitor Atencia
4 shared publications- Annette Möller
Faculty of Business Administration and Economics Bielefeld University Bielefeld Germany
4 shared publications Ben Hooper
4 shared publications- Bert Van Schaeybroeck
Royal Meteorological Institute of Belgium (RMIB), Brussels, Belgium
4 shared publications
Latest Publications
- Open Access
- 22 June 2026
- Published by Copernicus GmbH
AI-aided Convection Forecast for Aviation Planning
In the context of aviation planning over Europe, a two-day forecast of the convective activity is required. Such a product, named the Cross-Border Convection Forecast (CBCF), is regularly produced during the convective season by several meteorological services across Europe and coordinated by EUMETNET. Using a collaborative editing platform, human forecasters generate forecast polygons to quantify the expected convective situation. In particular, the forecast includes the expected degree of organization of the convective cells and the expected likelihood of occurrence.The goal of this work is to provide an automatic guidance product based on AI methods to assist the forecasters. Moreover, the guidance product should directly provide the final CBCF-forecast category, i.e. degree of organization and its likelihood of occurrence. We report on the development of such a guidance product, where the basic forecast data are taken from the ICON-EU ensemble forecasts that are operationally produced at the German Meteorological Service and cover the whole desired European forecast domain. Based on this ensemble data, several derived statistical quantities are computed and serve as input to the AI method.One significant challenge in the development of the guidance product is the definition of the target for training of the AI model since direct observations are not available for the final forecast categories. As a result, another quantity needs to be used as target for the training. In our approach we employ the area fraction of lightning, being a derived product of lightning observations and defined as the area covered by lightning within a radius of 50km. We show preliminary results for the AI methods that use the area fraction of lightning as target. Moreover, we discuss possible modifications of this quantity to adapt it better to the desired final product.
- 22 June 2026
- Published by Copernicus GmbH
Open AccessAI-aided Convection Forecast for Aviation Planning
In the context of aviation planning over Europe, a two-day forecast of the convective activity is required. Such a product, named the Cross-Border Convection Forecast (CBCF), is regularly produced during the convective season by several meteorological services across Europe and coordinated by EUMETNET. Using a collaborative editing platform, human forecasters generate forecast polygons to quantify the expected convective situation. In particular, the forecast includes the expected degree of organization of the convective cells and the expected likelihood of occurrence.The goal of this work is to provide an automatic guidance product based on AI methods to assist the forecasters. Moreover, the guidance product should directly provide the final CBCF-forecast category, i.e. degree of organization and its likelihood of occurrence. We report on the development of such a guidance product, where the basic forecast data are taken from the ICON-EU ensemble forecasts that are operationally produced at the German Meteorological Service and cover the whole desired European forecast domain. Based on this ensemble data, several derived statistical quantities are computed and serve as input to the AI method.One significant challenge in the development of the guidance product is the definition of the target for training of the AI model since direct observations are not available for the final forecast categories. As a result, another quantity needs to be used as target for the training. In our approach we employ the area fraction of lightning, being a derived product of lightning observations and defined as the area covered by lightning within a radius of 50km. We show preliminary results for the AI methods that use the area fraction of lightning as target. Moreover, we discuss possible modifications of this quantity to adapt it better to the desired final product.
- Open Access
- 22 June 2026
- Published by Copernicus GmbH
Towards User-Oriented Verification of Weather Warnings: Insights from Public and Stakeholder Preferences
Verifying the quality of weather warnings is important for improving warnings, building trust, and enabling stakeholders to choose the right course of action. But what makes a warning valuable from a user’s point of view? The recipients of weather warnings, such as emergency response teams, commercial actors, or the general public, may each prioritize different aspects of weather warnings. For instance, some may value the correct intensity of an event over its correct timing, while others may prefer correct location over the correct duration. In addition, the trade-off between over- and under-forecasted weather events may vary between stakeholders as some might prioritize minimizing missed events while others prefer not having too many false alarms. To better capture these individual preferences, this work showcases the efforts made at the German Meteorological Service (Deutscher Wetterdienst, DWD) in the context of the renewal of its warning system (i.e., the RainBoW program: “Risk-based, Application-oriented and Individualizable Provision of Optimized Warning Information”) to understand how users perceive the quality of warnings.To explore what properties of a warning are the most important to the general public - the largest group of recipients of weather warnings - we conducted two non-representative surveys during two ‘open house’ DWD events. Participants recorded how satisfied they were with forecasts if specific attributes such as intensity, timing, location, or persistence were incorrect and affected the outcome of their planned activities. In addition to this public survey, a workshop with stakeholders from disaster response teams provided further insights into their specific needs and interests towards forecast quality.With this information from different stakeholders, we aim to develop verification methods that take these preferences into account when evaluating the accuracy of a warning. The development of new verification methods may be supported by user reports of the current weather collected through the official DWD weather app.A key component of any user-oriented verification approach is an effective communication of the results. This requires translating complex statistical information into clear practical information that is accessible and meaningful for non-specialist audiences.
- 22 June 2026
- Published by Copernicus GmbH
Open AccessTowards User-Oriented Verification of Weather Warnings: Insights from Public and Stakeholder Preferences
Verifying the quality of weather warnings is important for improving warnings, building trust, and enabling stakeholders to choose the right course of action. But what makes a warning valuable from a user’s point of view? The recipients of weather warnings, such as emergency response teams, commercial actors, or the general public, may each prioritize different aspects of weather warnings. For instance, some may value the correct intensity of an event over its correct timing, while others may prefer correct location over the correct duration. In addition, the trade-off between over- and under-forecasted weather events may vary between stakeholders as some might prioritize minimizing missed events while others prefer not having too many false alarms. To better capture these individual preferences, this work showcases the efforts made at the German Meteorological Service (Deutscher Wetterdienst, DWD) in the context of the renewal of its warning system (i.e., the RainBoW program: “Risk-based, Application-oriented and Individualizable Provision of Optimized Warning Information”) to understand how users perceive the quality of warnings.To explore what properties of a warning are the most important to the general public - the largest group of recipients of weather warnings - we conducted two non-representative surveys during two ‘open house’ DWD events. Participants recorded how satisfied they were with forecasts if specific attributes such as intensity, timing, location, or persistence were incorrect and affected the outcome of their planned activities. In addition to this public survey, a workshop with stakeholders from disaster response teams provided further insights into their specific needs and interests towards forecast quality.With this information from different stakeholders, we aim to develop verification methods that take these preferences into account when evaluating the accuracy of a warning. The development of new verification methods may be supported by user reports of the current weather collected through the official DWD weather app.A key component of any user-oriented verification approach is an effective communication of the results. This requires translating complex statistical information into clear practical information that is accessible and meaningful for non-specialist audiences.
- 2 January 2026
- Published by Springer Nature
- p. 153-180
Comparison of Model Output Statistics and Neural Networks to Postprocess Wind Gusts
- 2 January 2026
- Published by Springer Nature
- p. 153-180
Comparison of Model Output Statistics and Neural Networks to Postprocess Wind Gusts
- 2 January 2026
- Published by Springer Nature
- p. 43-65
Forecasting Lightning Probabilities Derived from the Lightning Potential Index Using Neural Networks
- 2 January 2026
- Published by Springer Nature
- p. 43-65
Forecasting Lightning Probabilities Derived from the Lightning Potential Index Using Neural Networks
- Open Access
- 8 August 2025
- Published by Copernicus GmbH
Machine Learning Methods for the Postprocessing and Seamless Blending of Ensemble Forecasts
Numerical weather prediction (NWP) models can deliver high-resolution forecasts, but achieving optimal performance requires the fine-tuning of model parameters and adjusting the physical models for different spatial, temporal, and seasonal scopes. Models differ in forecast accuracy for specific lead times and seasonal environments, offering a variety of fragmented forecasts exhibiting temporal discontinuities. This happens with the two regional ICON variants for Germany, ICON-D2 and ICON-RUC, which are operational at DWD and differ in their physical models and forecast range. Although both operate on a 2 km scale, ICON-RUC focuses on convection, is more computationally expensive, and it provides shorter lead times (+14 hours for ICON-RUC vs. +48 hours for ICON-D2). Therefore, blending is necessary for a seamless forecast for up to 48 hours. Moreover, NWP systems are biased and postprocessing is required for calibration and error correction.As part of the SINFONY 3.0 project at DWD, our goal is to use machine learning methods to develop a postprocessing framework that delivers an improved combined probabilistic forecast that is calibrated and seamlessly blends the output data from both NWP models, with a focus on hourly precipitation. DWD’s radar network measurements is used as the ground truth for training.We build on previous work that utilizes ML to improve probabilistic forecasts: Grönquist et al. [1] used deep U-Nets for bias and spread estimation. Rempel et al. [2] focus on the blending of ensemble nowcasting and ensemble NWP, and also show that their network can improve calibration. Primo et al. [3] generate calibrated probabilistic distributions using a neural network and include additional contextual data features such as lead time, seasonal, and orographic parameters to improve predictions. This work presents a neural network that blends both NWP systems and provides a calibrated probability.[1] Peter Grönquist, Chengyuan Yao, Tal Ben-Nun, Nikoli Dryden, Peter Dueben, Shigang Li, and Torsten Hoefler. Deep learning for post-processing ensemble weather forecasts. Philos. Trans. A Math. Phys. Eng. Sci., 379(2194):20200092, April 2021.[2] Martin Rempel, Peter Schaumann, Reinhold Hess, Volker Schmidt, and Ulrich Blahak. Adaptive blending of probabilistic precipitation forecasts with emphasis on calibration and temporal forecast consistency. Artificial Intelligence for the Earth Systems, 1(4), October 2022.[3] Cristina Primo, Benedikt Schulz, Sebastian Lerch, and Reinhold Hess. Comparison of model output statistics and neural networks to postprocess wind gusts. arXiv [stat.AP], January 2024.
- 8 August 2025
- Published by Copernicus GmbH
Open AccessMachine Learning Methods for the Postprocessing and Seamless Blending of Ensemble Forecasts
Numerical weather prediction (NWP) models can deliver high-resolution forecasts, but achieving optimal performance requires the fine-tuning of model parameters and adjusting the physical models for different spatial, temporal, and seasonal scopes. Models differ in forecast accuracy for specific lead times and seasonal environments, offering a variety of fragmented forecasts exhibiting temporal discontinuities. This happens with the two regional ICON variants for Germany, ICON-D2 and ICON-RUC, which are operational at DWD and differ in their physical models and forecast range. Although both operate on a 2 km scale, ICON-RUC focuses on convection, is more computationally expensive, and it provides shorter lead times (+14 hours for ICON-RUC vs. +48 hours for ICON-D2). Therefore, blending is necessary for a seamless forecast for up to 48 hours. Moreover, NWP systems are biased and postprocessing is required for calibration and error correction.As part of the SINFONY 3.0 project at DWD, our goal is to use machine learning methods to develop a postprocessing framework that delivers an improved combined probabilistic forecast that is calibrated and seamlessly blends the output data from both NWP models, with a focus on hourly precipitation. DWD’s radar network measurements is used as the ground truth for training.We build on previous work that utilizes ML to improve probabilistic forecasts: Grönquist et al. [1] used deep U-Nets for bias and spread estimation. Rempel et al. [2] focus on the blending of ensemble nowcasting and ensemble NWP, and also show that their network can improve calibration. Primo et al. [3] generate calibrated probabilistic distributions using a neural network and include additional contextual data features such as lead time, seasonal, and orographic parameters to improve predictions. This work presents a neural network that blends both NWP systems and provides a calibrated probability.[1] Peter Grönquist, Chengyuan Yao, Tal Ben-Nun, Nikoli Dryden, Peter Dueben, Shigang Li, and Torsten Hoefler. Deep learning for post-processing ensemble weather forecasts. Philos. Trans. A Math. Phys. Eng. Sci., 379(2194):20200092, April 2021.[2] Martin Rempel, Peter Schaumann, Reinhold Hess, Volker Schmidt, and Ulrich Blahak. Adaptive blending of probabilistic precipitation forecasts with emphasis on calibration and temporal forecast consistency. Artificial Intelligence for the Earth Systems, 1(4), October 2022.[3] Cristina Primo, Benedikt Schulz, Sebastian Lerch, and Reinhold Hess. Comparison of model output statistics and neural networks to postprocess wind gusts. arXiv [stat.AP], January 2024.
- Open Access
- 16 July 2025
- Published by Copernicus GmbH
First Results and Evaluation of Automated Wind Gust Warnings
As part of the ongoing renewal of the warning system in the program RainBoW ("Risk-based, Application-oriented and INdividualizaBle Provision of Optimized Warning Information"), the German Meteorological Service (Deutscher Wetterdienst, DWD) aims to produce warning information for up to seven days to inform the public early-on about severe weather. Maintaining such a forecast horizon at a high update frequency requires automated processing chains. In this contribution, we present the current state of a running prototype for automated wind gust warnings.The prototype’s processing chain uses modern messaging techniques to combine ensemble wind gust forecasts from DWD's in-house model ICON in multiple setups (local area version, EU version, global version). The model data is subsequently translated into gridded threshold exceedance probabilities for different warning levels. If exceedance probabilities are high enough, the according warning level will be assigned. New incoming forecast data from an ICON model version is used to update the exceedance probabilities, which in turn may lead to changes in the assigned warning level for a fixed location and forecast date. We therefore subsequently apply a smoothing across updates to limit the amount of such changes. For completeness, the prototype also provides consistent information about the wind direction, about possible worst-case scenarios, represented by the ensemble maximum of wind gusts, and about uncertainties linked to the ICON model predictions as well as threshold exceedance probabilities. Wherever the spatial resolution of ICON input data is sufficiently high, exceedance thresholds are calculated across multiple topographic layers (i.e., elevation intervals). This approach helps better represent the vertical structure of warnings and reduces false alarms in areas near mountain peaks and valleys.Monitoring and verification tools were developed to compare (smoothed) gust warning fields with SYNOP observations for case studies of extreme weather situations as well as longer periods. Immediate forecast monitoring indicated that the spatial extent of resulting automated warning data is generally in good agreement with observations. Longer-term verification statistics showed that smoothing increased the probability of detection compared to the original automated warning results. These insights have helped us to improve the configuration of the wind gust prototype.
- 16 July 2025
- Published by Copernicus GmbH
Open AccessFirst Results and Evaluation of Automated Wind Gust Warnings
As part of the ongoing renewal of the warning system in the program RainBoW ("Risk-based, Application-oriented and INdividualizaBle Provision of Optimized Warning Information"), the German Meteorological Service (Deutscher Wetterdienst, DWD) aims to produce warning information for up to seven days to inform the public early-on about severe weather. Maintaining such a forecast horizon at a high update frequency requires automated processing chains. In this contribution, we present the current state of a running prototype for automated wind gust warnings.The prototype’s processing chain uses modern messaging techniques to combine ensemble wind gust forecasts from DWD's in-house model ICON in multiple setups (local area version, EU version, global version). The model data is subsequently translated into gridded threshold exceedance probabilities for different warning levels. If exceedance probabilities are high enough, the according warning level will be assigned. New incoming forecast data from an ICON model version is used to update the exceedance probabilities, which in turn may lead to changes in the assigned warning level for a fixed location and forecast date. We therefore subsequently apply a smoothing across updates to limit the amount of such changes. For completeness, the prototype also provides consistent information about the wind direction, about possible worst-case scenarios, represented by the ensemble maximum of wind gusts, and about uncertainties linked to the ICON model predictions as well as threshold exceedance probabilities. Wherever the spatial resolution of ICON input data is sufficiently high, exceedance thresholds are calculated across multiple topographic layers (i.e., elevation intervals). This approach helps better represent the vertical structure of warnings and reduces false alarms in areas near mountain peaks and valleys.Monitoring and verification tools were developed to compare (smoothed) gust warning fields with SYNOP observations for case studies of extreme weather situations as well as longer periods. Immediate forecast monitoring indicated that the spatial extent of resulting automated warning data is generally in good agreement with observations. Longer-term verification statistics showed that smoothing increased the probability of detection compared to the original automated warning results. These insights have helped us to improve the configuration of the wind gust prototype.
- Open Access
- 30 June 2025
- Published by Copernicus GmbH
Seamless global lightning forecasts based on convolutional graph neural networks
Global lightning forecasts are of particular interest to the aviation industry. However, global weather forecasts are typically based on numerical weather prediction model (NWP) runs with resolutions that cannot resolve deep convection, hence it needs to be parameterized. Although lightning may be diagnosed in these models, e.g. with the Lightning Potential Index (LPI,1), the quality of such forecasts is only mediocre. Successful attempts have been made to produce lightning probability forecasts with neural networks that use the LPI as input features and lighting observations as target (2,3). Moreover, high resolution global lightning data is available (GLD360, Vaisala, 4) which has been used for nowcasting of thunderstorms up to 2 hours lead time (5).This work shows how the combination of global lightning observations (GLD360 from Vaisala, 4) with NWP can significantly improve the global lightning probability forecast for up to 8 hours. A graph neural network initialized with global lightning observations of the past hours has been developed. Observations are transferred into a latent space by a fully connected neural network and then integrated forward in time. Multiple loops using a convolutional graph neural network are employed to include spatial information. Here, NWP forecasts are fed in as feature and are the driver for the initiation of new thunderstorms, their propagation and decay.The model is trained with one year ICON Reanalysis data and applied to global ICON ensemble forecasts. Preliminary verification results based on the resolution component of the Brier Score show that initializing the model with global lightning observations of the past hours significantly outperforms a model version without that initialization. This improvement remains for lead times up to 8 hours, then the verification scores of the two model versions converge. Furthermore, case studies show that the lightning is well resolved during the first forecast hours, i.e. beyond the typical 2 hour nowcasting horizon.References:(1) Schröder et al., 2022: Subgrid scale Lightning Potential Index for ICON with parameterized convection. Reports on ICON (10), DOI: https://doi.org/10.5676/dwd_pub/nwv/icon_010 .(2) Baumgartner, M., Schröder, G., and Primo, C.: Forecasting lightning probabilities derived from the Lightning Potential Index using neural networks, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-34, https://doi.org/10.5194/ems2023-34, 2023.(3) Schröder, G., Baumgartner, M., Primo, C., and Theis, S.: Challenges in training neural networks for global lightning probability forecasts, EMS Annual Meeting 2024, Barcelona, Spain, 1–6 Sep 2024, EMS2024-475, https://doi.org/10.5194/ems2024-475, 2024.(4) https://www.vaisala.com/en/products/systems/lightning/gld360(5) Müller, R., Barleben, A., Haussler, S., & Jerg, M. (2022). A Novel Approach for the Global Detection and Nowcasting of Deep Convection and Thunderstorms. Remote Sensing, 14(14), 3372. https://doi.org/10.3390/rs14143372
- 30 June 2025
- Published by Copernicus GmbH
Open AccessSeamless global lightning forecasts based on convolutional graph neural networks
Global lightning forecasts are of particular interest to the aviation industry. However, global weather forecasts are typically based on numerical weather prediction model (NWP) runs with resolutions that cannot resolve deep convection, hence it needs to be parameterized. Although lightning may be diagnosed in these models, e.g. with the Lightning Potential Index (LPI,1), the quality of such forecasts is only mediocre. Successful attempts have been made to produce lightning probability forecasts with neural networks that use the LPI as input features and lighting observations as target (2,3). Moreover, high resolution global lightning data is available (GLD360, Vaisala, 4) which has been used for nowcasting of thunderstorms up to 2 hours lead time (5).This work shows how the combination of global lightning observations (GLD360 from Vaisala, 4) with NWP can significantly improve the global lightning probability forecast for up to 8 hours. A graph neural network initialized with global lightning observations of the past hours has been developed. Observations are transferred into a latent space by a fully connected neural network and then integrated forward in time. Multiple loops using a convolutional graph neural network are employed to include spatial information. Here, NWP forecasts are fed in as feature and are the driver for the initiation of new thunderstorms, their propagation and decay.The model is trained with one year ICON Reanalysis data and applied to global ICON ensemble forecasts. Preliminary verification results based on the resolution component of the Brier Score show that initializing the model with global lightning observations of the past hours significantly outperforms a model version without that initialization. This improvement remains for lead times up to 8 hours, then the verification scores of the two model versions converge. Furthermore, case studies show that the lightning is well resolved during the first forecast hours, i.e. beyond the typical 2 hour nowcasting horizon.References:(1) Schröder et al., 2022: Subgrid scale Lightning Potential Index for ICON with parameterized convection. Reports on ICON (10), DOI: https://doi.org/10.5676/dwd_pub/nwv/icon_010 .(2) Baumgartner, M., Schröder, G., and Primo, C.: Forecasting lightning probabilities derived from the Lightning Potential Index using neural networks, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-34, https://doi.org/10.5194/ems2023-34, 2023.(3) Schröder, G., Baumgartner, M., Primo, C., and Theis, S.: Challenges in training neural networks for global lightning probability forecasts, EMS Annual Meeting 2024, Barcelona, Spain, 1–6 Sep 2024, EMS2024-475, https://doi.org/10.5194/ems2024-475, 2024.(4) https://www.vaisala.com/en/products/systems/lightning/gld360(5) Müller, R., Barleben, A., Haussler, S., & Jerg, M. (2022). A Novel Approach for the Global Detection and Nowcasting of Deep Convection and Thunderstorms. Remote Sensing, 14(14), 3372. https://doi.org/10.3390/rs14143372
- Open Access
- 30 June 2025
- Published by Copernicus GmbH
A User-oriented approach to verifying weather warnings
Verifying the quality of weather warnings is crucial to improving warnings, building trust, and enabling stakeholders to choose the right course of action in the face of a warning. In the context of the renewal of the warning system at Deutscher Wetterdienst (i.e., the RainBoW program: “Risk-based, Application-oriented and Individualizable Provision of Optimized Warning Information”) with the main goal to align weather warnings more strongly with the needs of their users, there is a need for verification approaches which are specifically user-oriented.The recipients of weather warnings, such as emergency response teams, commercial actors, or the general public, however, have different preferences as to which aspects of weather warnings are the most important to them. This could be valuing the correct intensity of an event over its correct timing in a forecast, or the correct location over the correct duration. It could also mean that the trade-off between over- and under-forecasted weather events could vary between stakeholders as some might prioritize minimizing missed events (i.e., increasing the hit rate, POD) while others prefer not having too many false alarms (i.e., reducing the False Alarm Rate, FAR). Incorporating these user-specific requirements provides the potential to build on common verification methods and expanding them, making them more applicable and useful to individual stakeholders.Additionally, communicating the results in a way that recipients understand and benefit from information on forecast quality is an important step in a user-oriented verification approach. This likely requires translating complex statistical information into real-world examples that are comprehensible for non-academic stakeholders as well.Therefore, this work showcases the efforts made at Deutscher Wetterdienst to provide user-oriented weather warning verification byconducting small surveys to find out which aspects of weather warnings are more important to a certain user group, testing verification methods that take these preferences into account when evaluating the accuracy of a warning, developing and testing comprehensible and user-focused ways of communicating the quality of warnings to user-groups.
- 30 June 2025
- Published by Copernicus GmbH
Open AccessA User-oriented approach to verifying weather warnings
Verifying the quality of weather warnings is crucial to improving warnings, building trust, and enabling stakeholders to choose the right course of action in the face of a warning. In the context of the renewal of the warning system at Deutscher Wetterdienst (i.e., the RainBoW program: “Risk-based, Application-oriented and Individualizable Provision of Optimized Warning Information”) with the main goal to align weather warnings more strongly with the needs of their users, there is a need for verification approaches which are specifically user-oriented.The recipients of weather warnings, such as emergency response teams, commercial actors, or the general public, however, have different preferences as to which aspects of weather warnings are the most important to them. This could be valuing the correct intensity of an event over its correct timing in a forecast, or the correct location over the correct duration. It could also mean that the trade-off between over- and under-forecasted weather events could vary between stakeholders as some might prioritize minimizing missed events (i.e., increasing the hit rate, POD) while others prefer not having too many false alarms (i.e., reducing the False Alarm Rate, FAR). Incorporating these user-specific requirements provides the potential to build on common verification methods and expanding them, making them more applicable and useful to individual stakeholders.Additionally, communicating the results in a way that recipients understand and benefit from information on forecast quality is an important step in a user-oriented verification approach. This likely requires translating complex statistical information into real-world examples that are comprehensible for non-academic stakeholders as well.Therefore, this work showcases the efforts made at Deutscher Wetterdienst to provide user-oriented weather warning verification byconducting small surveys to find out which aspects of weather warnings are more important to a certain user group, testing verification methods that take these preferences into account when evaluating the accuracy of a warning, developing and testing comprehensible and user-focused ways of communicating the quality of warnings to user-groups.
- Open Access
- 16 August 2024
- Published by Copernicus GmbH
Challenges in training neural networks for global lightning probability forecasts
Forecasting lightning is generally challenging for numerical weather models with parameterized convection. Nonetheless, for the aviation industry reliable lightning forecasts are needed for larger domains for which running convection permitting models is not feasible yet. For models with coarser resolution, parameterizations exist that derive the potential for lightning based on the model's convection scheme. One example is the subgrid-scale lightning potential index (LPI, Schröder et al. 2022). However, this index inherits model biases such as a biased onset of convection. Furthermore, for ensemble prediction systems, the spread of the LPI is not calibrated either. Baumgartner et al. (2023) have shown that applying a neural network within the European domain can correct model errors like the wrong timing as well as the ensemble spread. Nevertheless, applying this approach to the global domain leads to further challenges, e.g. the different behaviour of the numerical model in different climatological regions of the world. From a technical point of view, a global product needs to be trained on significantly larger datasets. Furthermore, after an update of the numerical model usually only a few months of data are available for (re-)training which is inherently a biased dataset. In this work it is shown how to deal with these challenges by using the global ensemble prediction system of ICON. The global lightning data of Vaisala (GLD360, 3) serves as ground truth. A key component of the neural network is a climate feature that allows the neural network to differentiate between the different climatological regions. This is accompanied by features like solar elevation and local time. The benefits of these new features/predictors are shown using the Brier score. The diurnal cycle, as well as region dependent climatological biases, are corrected. It is also shown that even without the availability of training data for a whole year, the neural network can learn corrections e.g. from the northern hemisphere summer and apply them to the southern hemisphere summer and vice versa.References:(1) Schröder et al., 2022: Subgrid scale Lightning Potential Index for ICON with parameterized convection. Reports on ICON (10), DOI: https://doi.org/10.5676/dwd_pub/nwv/icon_010 .(2) Baumgartner, M., Schröder, G., and Primo, C.: Forecasting lightning probabilities derived from the Lightning Potential Index using neural networks, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-34, https://doi.org/10.5194/ems2023-34, 2023.(3) https://www.vaisala.com/en/products/systems/lightning/gld3
- 16 August 2024
- Published by Copernicus GmbH
Open AccessChallenges in training neural networks for global lightning probability forecasts
Forecasting lightning is generally challenging for numerical weather models with parameterized convection. Nonetheless, for the aviation industry reliable lightning forecasts are needed for larger domains for which running convection permitting models is not feasible yet. For models with coarser resolution, parameterizations exist that derive the potential for lightning based on the model's convection scheme. One example is the subgrid-scale lightning potential index (LPI, Schröder et al. 2022). However, this index inherits model biases such as a biased onset of convection. Furthermore, for ensemble prediction systems, the spread of the LPI is not calibrated either. Baumgartner et al. (2023) have shown that applying a neural network within the European domain can correct model errors like the wrong timing as well as the ensemble spread. Nevertheless, applying this approach to the global domain leads to further challenges, e.g. the different behaviour of the numerical model in different climatological regions of the world. From a technical point of view, a global product needs to be trained on significantly larger datasets. Furthermore, after an update of the numerical model usually only a few months of data are available for (re-)training which is inherently a biased dataset. In this work it is shown how to deal with these challenges by using the global ensemble prediction system of ICON. The global lightning data of Vaisala (GLD360, 3) serves as ground truth. A key component of the neural network is a climate feature that allows the neural network to differentiate between the different climatological regions. This is accompanied by features like solar elevation and local time. The benefits of these new features/predictors are shown using the Brier score. The diurnal cycle, as well as region dependent climatological biases, are corrected. It is also shown that even without the availability of training data for a whole year, the neural network can learn corrections e.g. from the northern hemisphere summer and apply them to the southern hemisphere summer and vice versa.References:(1) Schröder et al., 2022: Subgrid scale Lightning Potential Index for ICON with parameterized convection. Reports on ICON (10), DOI: https://doi.org/10.5676/dwd_pub/nwv/icon_010 .(2) Baumgartner, M., Schröder, G., and Primo, C.: Forecasting lightning probabilities derived from the Lightning Potential Index using neural networks, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-34, https://doi.org/10.5194/ems2023-34, 2023.(3) https://www.vaisala.com/en/products/systems/lightning/gld3
- Open Access
- 6 July 2023
- Published by Copernicus GmbH
Forecasting lightning probabilities derived from the Lightning Potential Index using neural networks
Forecasting the occurrence of thunderstorms is a well-known challenge in weather prediction. Since a thunderstorm is by definition accompanied by at least one lightning, we aim to forecast the occurrence of at least one lightning within a pre-defined area and a pre-defined time-interval. Numerous prior research studies on forecasting and detecting lightnings provide a rich database of model diagnostics and observational data. One example of such a model diagnostic is the so-called (subgrid-scale) Lightning Potential Index (LPI), that was recently implemented in the operational ICON-model (1) and is now available operationally, even in the ICON-EU ensemble. On the other hand, there are extensive observation networks that provide lightning observations, such as the Linet-network (2) that provides lightning observations over Europe. In our work, we use lightning observations from the Linet-network as ground truth and establish a translation from LPI to lightning probabilities. For this task, we trained neural networks to predict the desired lightning probabilities in a "global postprocessing mode", i.e. using the same network for the forecasts on the whole domain of ICON-EU, which is significantly larger than the domain of the Linet-network. We present the setup of the postprocessing method together with details of the training of the neural networks and show first results from its forecasts and their evaluation. In particular, these lightning probabilities outperform raw model probabilities as derived by counting the exceedance of a LPI-threshold and thereby avoid the need to even define such a threshold. We will also address aspects of producing stable operational forecasts by applying an ensemble of neural networks. References: (1) Schröder et al., 2022: Subgrid scale Lightning Potential Index for ICON with parameterized convection. Reports on ICON (10), DOI: 10.5676/DWD_pub/nwv/icon_010 (2) Betz et al., 2009: LINET - an international lightning detection network in europe. Atmos. Res., 91, 564–573, DOI: 10.1016/j.atmosres.2008.06.012
- 6 July 2023
- Published by Copernicus GmbH
Open AccessForecasting lightning probabilities derived from the Lightning Potential Index using neural networks
Forecasting the occurrence of thunderstorms is a well-known challenge in weather prediction. Since a thunderstorm is by definition accompanied by at least one lightning, we aim to forecast the occurrence of at least one lightning within a pre-defined area and a pre-defined time-interval. Numerous prior research studies on forecasting and detecting lightnings provide a rich database of model diagnostics and observational data. One example of such a model diagnostic is the so-called (subgrid-scale) Lightning Potential Index (LPI), that was recently implemented in the operational ICON-model (1) and is now available operationally, even in the ICON-EU ensemble. On the other hand, there are extensive observation networks that provide lightning observations, such as the Linet-network (2) that provides lightning observations over Europe. In our work, we use lightning observations from the Linet-network as ground truth and establish a translation from LPI to lightning probabilities. For this task, we trained neural networks to predict the desired lightning probabilities in a "global postprocessing mode", i.e. using the same network for the forecasts on the whole domain of ICON-EU, which is significantly larger than the domain of the Linet-network. We present the setup of the postprocessing method together with details of the training of the neural networks and show first results from its forecasts and their evaluation. In particular, these lightning probabilities outperform raw model probabilities as derived by counting the exceedance of a LPI-threshold and thereby avoid the need to even define such a threshold. We will also address aspects of producing stable operational forecasts by applying an ensemble of neural networks. References: (1) Schröder et al., 2022: Subgrid scale Lightning Potential Index for ICON with parameterized convection. Reports on ICON (10), DOI: 10.5676/DWD_pub/nwv/icon_010 (2) Betz et al., 2009: LINET - an international lightning detection network in europe. Atmos. Res., 91, 564–573, DOI: 10.1016/j.atmosres.2008.06.012