Statistical post-processing of operational dual-resolution wind-speed ensemble forecasts
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Statistical post-processing of operational dual-resolution wind-speed ensemble forecasts".
Jane: Here's the gist of the paper. Statistical post-processing of operational dual-resolution wind-speed ensemble forecasts provides a detailed comparative analysis of forecast skill for raw and post-processed medium- and extended-range wind-speed ensemble forecasts at 9 km and 36 km horizontal resolutions, revealing that all post-processed forecasts outperform the raw ensemble predictions in terms of probabilistic calibration and point forecast accuracy.
Tom: Study Objective This study provides a detailed comparative analysis of the forecast skill of the raw and post-processed medium- and extended-range wind-speed ensemble forecasts of the European Centre for Medium-Range Weather Forecasts issued at 9 km and 36 km horizontal resolutions, respectively, and their various mixtures.
Jane: The aim of this study is to compare raw and post-processed ECMWF operational dual-resolution 10-m wind-speed forecasts and investigate the impact of incorporating varying numbers of higher-resolution members into the calibration process.
Tom: Data Description The dataset at hand comprises 50-member medium-range forecasts of 10-m wind-speed at TCO1279 resolution and 100-member extended-range forecasts at TCO319 resolution, obtained from perturbed initial conditions and/or parametrizations for 8726 synoptic observation (SYNOP) stations from July 1, 2023, to May 31, 2024.
Paper summary: Jane: So, to summarize what this paper, "Statistical post-processing of operational dual-resolution wind-speed ensemble forecasts," is about, it’s a comparative analysis of raw and post-processed medium and extended range wind speed ensemble forecasts from the European Centre for Medium Range Weather Forecasts.
Tom: The thesis they are pushing is that statistical post processing, specifically using an ensemble model output statistic approach for calibration, improves the forecast skill in terms of probabilistic calibration and point forecast accuracy.
Lu: They investigate how mixing different resolutions—specifically nine kilometer and thirty-six kilometer horizontal resolutions—and varying ensemble sizes changes the performance when you look at these wind speed predictions <ref:2506.15578#pg1>.
Meng: The data they use is quite substantial; they are looking at fifty-member medium range forecasts and one hundred-member extended range forecasts, all from a set of stations <ref:2506.15578#pg1,and 100-member extended range>.
Lalam: The study uses three different spatial training data selection techniques to test how well those models calibrate against the training data.
Jane: What matters is that they find that incorporating high resolution members into low resolution ensemble forecasts yields the most significant gains when the highest number of high resolution members are included in the calibration process.
Tom: They also look at how spatial resolution and ensemble size interact, showing that spatial resolution is superior to ensemble size alone for medium range temperature predictions, but it depends on what you're forecasting.
Lu: The paper shows that statistical post-processing significantly reduces the performance differences between various single and dual resolution configurations when they are combined.
Meng: From a production standpoint, this means if you have a lower resolution forecast available, applying these calibration techniques could bring its reliability up to match or exceed what you'd get from a higher resolution model.
Lalam: They also note that the local EMOS model provides the most substantial overall improvement relative to the raw forecasts in terms of all studied evaluation metrics.
Jane: Essentially, they are showing that post-processing is a powerful tool for refining ensemble weather forecasts, especially when you're dealing with dual resolution data.
Tom: It really seems like the key mechanism here is that the statistical distribution correction helps bridge the gap between what the model predicts and what actually happens in terms of probabilistic accuracy.
Lu: The paper establishes that while dual resolution forecasts are superior in certain areas, post-processing makes them more robust overall across all evaluation metrics studied.
Meng: So, for us at scale, this suggests that if we can automate this statistical post-processing step efficiently, it could be a way to gain performance without needing to run the most computationally expensive models constantly.
Lalam: The results indicate that the dual resolution forecast is significantly superior to single resolution predictions only when you look at the mean quantile scores for the low and very high percentiles.
Conclusion: Tom: So, looking at "Statistical post-processing of operational dual-resolution wind-speed ensemble forecasts," what are the big takeaways from this work? The authors are Sándor Baran and Mária Lakatos.
Jane: The main point is that post processing—especially using the local and semi-local EMOS models—offers substantial improvements in forecast skill across all studied evaluation metrics when comparing raw ECMWF ensemble forecasts.
Lu: They found that the most significant gains come from configurations with the highest number of high resolution members, though even minimal inclusion still yields slight improvements for those specific settings.
Meng: In practical terms, this means if you are running a lower resolution forecast, applying this post-processing step can give you a much more reliable result in terms of calibration and accuracy.
Lalam: The paper concludes that the dual resolution forecast is significantly superior to single resolution predictions only when you look at the mean quantile scores for the low and very high percentiles.
Tom: It seems like the authors are arguing that while they can't always achieve perfect skill, using these statistical post-processing techniques gives you a solid performance boost compared to just running raw forecasts.
Jane: They are showing that this method is a tangible way to enhance the reliability of operational wind speed predictions by bridging the gap between model output and actual observed outcomes.
Lu: It points toward future work exploring more advanced methods like neural network based post-processing or multivariate approaches, which could be useful for understanding how different resolution members interact.
Meng: For us, the implication is that we need to consider this kind of calibration as a necessary step in our pipeline if we want to maximize the utility of our ensemble forecasts.
Lalam: The paper confirms that statistical post processing is a key technique for refining ensemble forecasts in this field.
Faculty of Informatics, University of Debrecen
stat.AP
Submitted: 2025-06-18
Updated: 2025-06-18
Comments: 26 pages, 17 figures
Journal ref: Quarterly Journal of the Royal Meteorological Society 152 (2026), no. 779, paper e70201
DOI: 10.1002/qj.70201
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 77/100
The gist: Statistical post-processing of operational dual-resolution wind-speed ensemble forecasts provides a detailed comparative analysis of forecast skill for raw and post-processed medium- and…
Key concepts
- Ensemble Forecasts
- These are multiple weather model runs starting from slightly different initial conditions to represent a range of possible outcomes. In this study, they were used for wind speed predictions at two different horizontal resolutions (9 km and 36 km).
- Post-Processing
- This involves using statistical methods to refine the raw ensemble output. The researchers tested various techniques, such as EMOS models, to improve the reliability of the forecasts by applying appropriate probability distributions.
- Probabilistic Calibration
- This measures how well a forecast's predicted probabilities match what actually happens. Post-processing methods were shown to significantly enhance this calibration compared to using the raw ensemble predictions alone.
- Spatial Resolution
- This refers to the level of detail in the weather forecast grid, measured in kilometers (e.g., 9 km vs. 36 km). The study investigated how mixing forecasts from different resolutions affects overall skill.
Terminology
Summary
Statistical post-processing of operational dual-resolution wind-speed ensemble forecasts provides a detailed comparative analysis of forecast skill for raw and post-processed medium- and extended-range wind-speed ensemble forecasts at 9 km and 36 km horizontal resolutions, revealing that all post-processed forecasts outperform the raw ensemble predictions in terms of probabilistic calibration and point forecast accuracy. p. 0
Study Objective
This study provides a detailed comparative analysis of the forecast skill of the raw and post-processed medium- and extended-range wind-speed ensemble forecasts of the European Centre for Medium-Range Weather Forecasts issued at 9 km and 36 km horizontal resolutions, respectively, and their various mixtures. The aim of this study is to compare raw and post-processed ECMWF operational dual-resolution 10-m wind-speed forecasts and investigate the impact of incorporating varying numbers of higher-resolution members into the calibration process. p. 2
Data Description
The dataset at hand comprises 50-member medium-range forecasts of 10-m wind-speed at TCO1279 resolution and 100-member extended-range forecasts at TCO319 resolution, obtained from perturbed initial conditions and/or parametrizations for 8726 synoptic observation (SYNOP) stations from July 1, 2023, to May 31, 2024. All forecasts are initialized at 00 UTC and the medium-range forecasts are limited to a 15-day horizon. p. 3
Calibration Methodology
The study utilizes the ensemble model output statistic approach for forecast calibration with three different spatial training data selection techniques. The EMOS method addresses the shortcomings of the raw predictions by applying a single parametric distribution to the ensemble outputs. p. 3 For wind-speed, a truncated normal (Thorarinsdottir and Gneiting, 2010) or a log-normal (Baran and Lerch, 2015) distribution is considered a more suitable choice. p. 3 The parameters of the EMOS predictive distribution are estimated by minimizing the mean value of a proper scoring rule over a carefully chosen training dataset. p. 4
Key Findings on Resolution and Ensemble Size
The study shows that spatial resolution is superior to the ensemble size; augmenting a sufficiently large ensemble of high-resolution forecasts with low-resolution predictions does not necessarily result in a gain in forecast skill. However, incorporating high-resolution members into low-resolution ensemble forecasts yields the most significant gains when the highest number of high-resolution members are included. p. 0 The local EMOS model provides the most substantial overall improvement relative to the raw forecasts, especially during the early forecast days. p. 8
Post-Processing Performance
The study demonstrates that all post-processed forecasts outperform the raw ensemble predictions in terms of probabilistic calibration and point forecast accuracy. The local and semi-local EMOS models provide the most substantial overall improvement relative to the raw forecasts. p. 8 For short lead times, the use of EMOS post-processed dual-resolution forecasts seems to be advantageous; it is significantly superior to the single-resolution setups in terms of the mean CRPS and mean QS for percentiles not less than the median. p. 9
Future Research Directions
The results suggest that one can investigate the performance of more sophisticated state-of-the-art post-processing methods such as the parametric machine-learningbased DRN (Rasp and Lerch, 2018) or versions of the non-parametric quantile regression. The study of multivariate post-processing approaches – classical two-step methods such as the ensemble copula coupling or data-driven techniques such as generative adversarial networks can also lead us to a better understanding of the effects of mixing high- and low-resolution ensemble members. p. 2 The increasing popularity of data-driven weather forecasts and the launch of the ensemble version of the ECMWF’s Artificial Intelligence Forecasting System currently issued at a 28 km grid resolution naturally induces the question of whether mixing IFS and AIFS-CRPS ensemble predictions results in improved predictive performance. p. 3
Conclusion
The study demonstrates that among the raw ECMWF ensemble forecasts, the low-resolution (100,0) pure low-resolution forecast exhibits the lowest predictive skill. Post-processing – particularly with local and semi-local EMOS models – offers substantial improvements in forecast skill in terms of all studied evaluation metrics. p. 8 The most significant gains are observed in configurations with the highest number of high-resolution members, though even minimal inclusion yields slight improvements. p. 0 Post-processing – particularly with local and semi-local EMOS models – offers substantial improvements in forecast skill in terms of all studied evaluation metrics, but the Brier scores for medium and high thresholds. p. 8 The local method demonstrates the best overall performance, especially during the early forecast days. p. 9 The dual-resolution forecast is significantly superior to the single-resolution predictions only in terms of the mean quantile scores for the low and very high percentiles. p. 9
References
Baran, S. and Lakatos, M. p. 0 (2024). Clustering-based spatial interpolation of parametric postprocessing models. p. 7
Baran, S. and Lerch, S. p. 3 p. 3 (2015). Log-normal distribution based emos models for probabilistic wind speed forecasting. p. 3
Baran, S., Leutbecher, M. p. 1 p. 1 p. 1, Szab´o, M. p. 2 p. 2, and Ben Bouall´egue, Z. p. 1 p. 1 p. 1 (2019). Statistical postprocessing of dual-resolution ensemble forecasts. p. 0
Baran, S., Szokol, P., and Szab´o, M. p. 2 (2021). Truncated generalized extreme value distribution-based ensemble model output statistics model for calibration of wind speed ensemble forecasts. p. 0
Ben Bouall´egue, Z., Haiden, T., Weber, N. J., Hamill, T. M., and Richardson, D. S. (2020). Accounting for representativeness in the verification of ensemble precipitation forecasts. p. 8
Bentzien, S. p. 5 and Friederichs, P. p. 5 (2014). Decomposition and graphical portrayal of the quantile score. p. 5
Bremnes, J. B. (2020). Ensemble postprocessing using quantile function regression based on neural networks and Bernstein polynomials. p. 3
Buizza, R. (2018). Ensemble forecasting and the need for calibration. p. 0
Chen, J., Janke, T., Steinke, F., and Lerch, S. p. 3 p. 3 (2024). Generative machine learning methods for multivariate ensemble postprocessing. p. 3
Clark, M., Gangopadhyay, S., Hay, L., Rajagopalan, B., and Wilby, R. (2004). The schaake shuffle: A method for reconstructing space–time variability in forecasted precipitation and temperature fields.
Dai, Y. and Hemri, S. p. 6 (2021). Spatially coherent postprocessing of cloud cover ensemble forecasts. p. 0
ECMWF (2024). p. 1 IFS Documentation CY49R1 – Part V: Ensemble Prediction System. p. 1
Gasc´on, E. p. 2, Lavers, D., Hamill, T. M., Richardson, D. S., Ben Bouall´egue, Z. p. 1, Leutbecher, M., and Pappenberger, F. (2019). Statistical postprocessing of dual-resolution ensemble precipitation forecasts across Europe. p. 2
Gneiting, T. p. 1 p. 1 p. 1 (2011). Making and evaluating point forecasts. p. 0
Gneiting, T. and Raftery, A. p. 4 E. (2007). Strictly proper scoring rules, prediction, and estimation. p. 4
Gneiting, T., Raftery, A. p. 4 E., Westveld, A. H., and Goldman, T. (2005). Calibrated probabilistic forecasting using ensemble model output statistics and minimum CRPS estimation. p. 0
Hemri, S., Scheuerer, M., Pappenberger, F., Bogner, K., and Haiden, T. (2014). Trends in the predictive performance of raw ensemble weather forecasts. p. 7
Jordan, A., Kr¨uger, F., and Lerch, S. (2019). Evaluating probabilistic forecasts with scoringRules. p. 0
Lang, S., Alexe, M., Clare, M. C. A., Roberts, C., Adewoyin, R., Bouall´egue, Z. p. 1 B., Chantry, M., Dramsch, J., Dueben, P. D., Hahner, S., Maciel, P., Prieto-Nemesio, A., O’Brien, C., Pinault, F., Polster, J., Raoult, B., Tietsche, S. and Leutbecher (2024). p. 1 AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the continuous ranked probability score. p. 4
Lerch, S. and Baran, S. p. 0 (2017). Similarity-based semilocal estimation of post-processing models. p. 0
Leutbecher, M. and Ben Bouall´egue, Z. (2020). On the probabilistic skill of dual-resolution ensemble forecasts. p. 0
Murphy, A. H. (1973). Hedging and skill scores for probability forecasts. p. 5
Politis, D. p. 6 N. and Romano, J. p. 6 P. (1994). The stationary bootstrap. p. 6
Rasp, S. and Lerch, S. (2018). Neural networks for postprocessing ensemble weather forecasts. p. 3
Schefzik, R., Thorarinsdottir, T. p. 3 p. 3 L., and Gneiting, T. p. 1 p. 1 (2013). Uncertainty quantification in complex simulation models using ensemble copula coupling. p. 1
Song, M., Yang, D., Lerch, S. p. 3 p. 3, Xia, X., Yagli, G. M., Bright, J. M., Shen, Y., Liu, B., and Liu (2024). p. 1 Non-crossing quantile regression neural network as a calibration tool for ensemble weather forecasts. p. 3
Szab´o, M., Gasc´on, E. p. 2, and Baran, S. p. 0 (2023). Parametric postprocessing of dual-resolution precipitation forecasts. p. 2
Taillardat, M. p. 3 (2021). Skewed and mixture of Gaussian distributions for ensemble postprocessing. p. 6
Thorarinsdottir, T. L. and Gneiting, T. (2010). Probabilistic forecasts of wind speed: Ensemble model output statistics by using heteroscedastic censored regression. p. 0
Vannitsem, S. p. 3, Bremnes, J. B., Demaeyer, J., Evans, G. R., Flowerdew, J., Hemri, S. p. 6, Lerch, S., Roberts, N., Theis, S., Atencia, A., Ben Boual´egue, Z. p. 1, Bhend, J., Dabernigig M. (2021). Statistical postprocessing for weather forecasts – review. p. 0
Veldkamp, S. p. 5, Whan, K., Dirksen, S., and Schmeits, M. (2021). Statistical postprocessing of wind speed forecasts using convolutional neural networks. p. 3
Wilks, D. S. (2019). Statistical Methods in the Atmospheric Sciences. p. 2
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Post-processing using local EMOS models can be employed to reduce deviations of various combinations, as
post-processing substantially reduces the deviations of the various combinations
andthe local model outperforms during the first six forecast days.
This allows AI systems to generate more reliable probabilistic forecasts by calibrating raw ensemble outputs based on spatial training data selection. -
The study confirms that incorporating high-resolution members into low-resolution ensemble forecasts can improve performance, as "the clear benefit of the other direction, namely, incorporating high-resolution members into low-resolution ensemble forecasts, where the most significant gains are observed in configurations with the highest number of high-resolution members." This suggests AI systems should prioritize augmenting lower resolution ensemble predictions with higher resolution members when aiming for improved skill.
-
The system can be optimized by training models using spatial selection strategies like regional, local, or semi-local EMOS approaches, where
the final configuration — 90 clusters and a 60-day training period — was selected by minimizing the mean CRPS over this validation period.
This enables the AI to dynamically select the optimal calibration method based on geographic data characteristics. -
For wind-speed forecasting, using a truncated normal distribution with parameters derived from ensemble statistics, such as
µ = a + bH H + bL f L and σ squared = c squared + d 2S squared,
provides a statistically sound framework for calibration. This ensures that the probabilistic forecasts accurately capture the non-negative and skewed nature of wind-speed data. -
The AI can improve its performance by utilizing quantile skill scores, as these scores reveal where improvements are most pronounced, noting that
the improvement is highest at the most extreme quantiles
for certain ensemble mixtures. This allows for targeted refinement of forecasts based on specific uncertainty levels.
Sources
Related papers
- Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization
- Strategic Causal Policy Learning with Manipulable Eligibility
- BayesJudge: Uncertainty-Aware Bayesian Meta-Evaluation of Human and LLM Judgments
- Automating structural reliability analysis with a multi-agent large language model framework
- Improving Minority Population Sampling with BISG Probabilities: Evidence from a Survey of Jewish Americans
- When Do Sports Leagues Stop Being Competitive? Bayesian Detection of Competitive-Balance Eras from Match-Level Data