[Paper Review] Post-processing output from ensembles with and without parametrised convection, to create accurate, blended, high-fidelity rainfall forecasts

Estíbaliz Gascón, A. Montani+1 more|arXiv (Cornell University)|Jan 11, 2023

Meteorological Phenomena and Simulations5 citations

TL;DR

This paper proposes a novel blending framework that combines post-processed global (ECMWF) and convection-resolving limited-area (COSMO-2I-EPS) ensemble rainfall forecasts using lead-time-weighted merging. By applying ecPoint-rainfall post-processing to the global ensemble and scale-selective neighborhood techniques to the LAM ensemble, the method produces more accurate, high-fidelity probabilistic 6-hour rainfall forecasts, with the blended product outperforming all individual components in verification—especially for summer convective events.

ABSTRACT

Flash flooding is a significant societal problem, but related precipitation forecasts are often poor. To address this, one can try to use output from convection-parametrising (global) ensembles, post-processed to forecast at point-scale, or convection-resolving limited area ensembles. In this study, we combine both. First, we apply the "ecPoint-rainfall" post-processing to the ECMWF global ensemble. Then, we use 2.2km COSMO LAM ensemble output (centred on Italy), and also post-process it using a scale-selective neighbourhood approach to compensate for insufficient members. The two components then undergo lead-time-weighted blending, to create the final probabilistic 6h rainfall forecasts. Product creation for forecasters constituted the "Italy Flash Flood use case" within the EU-funded MISTRAL project and it will be a real-time open-access product. One year of verification shows that ecPoint is the most skilful ensemble product. The post-processed COSMO ensemble adds most value to summer convective events in the evening, when the global model has an underprediction bias. In two heavy rainfall case studies we observed underestimation of the largest point totals in the raw ECMWF ensemble, and overestimation in the raw COSMO ensemble. However, ecPoint increase the value and highlighted best the most affected areas, whilst post-processing of COSMO diminished extremes by eradicating unreliable detail. The final merged products looked best from a user perspective and seemed to be the most skilful of all. Although our LAM post-processing does not implicitly include bias correction (a topic for further work) our study nonetheless provides a unique blueprint for successfully combining ensemble rainfall forecasts from global and LAM systems around the world. It also has important implications for forecast products as global ensembles move ever closer to having convection-permitting resolution.

Motivation & Objective

• Address the challenge of inaccurate point-scale rainfall forecasts for flash flood events, particularly due to sub-grid variability in global models.
• Improve forecast skill for convective precipitation by combining outputs from convection-parametrizing global ensembles and convection-resolving limited-area models (LAMs).
• Develop a seamless, high-fidelity probabilistic rainfall forecast system that maintains local detail while reducing noise and bias in high-resolution LAM outputs.
• Create a practical, real-time operational framework for flash flood warning systems, demonstrated in the MISTRAL project’s Italy Flash Flood use case.

Proposed method

• Apply ecPoint-rainfall post-processing to the ECMWF global ensemble to convert grid-box rainfall totals into accurate point-scale forecasts.
• Use a scale-selective neighborhood technique to post-process the 2.2 km COSMO-LAM ensemble, reducing spurious small-scale features and preserving reliable local details.
• Blend the two post-processed components using lead-time-weighted averaging to produce a final, seamless probabilistic 6-hour rainfall forecast.
• Validate the method using one year of real-time verification data and two detailed case studies of heavy rainfall events.
• Assess forecast skill using verification metrics, with a focus on extreme rainfall totals and spatial accuracy.
• Use visual and quantitative evaluation to assess user-relevance, credibility, and reliability of the final blended product.

Experimental results

Research questions

• RQ1 Can blending post-processed global and LAM ensemble forecasts improve probabilistic 6-hour point rainfall forecasts compared to raw or individually post-processed outputs?
• RQ2 How does the scale-selective neighborhood technique enhance the reliability of high-resolution LAM forecasts with limited ensemble members?
• RQ3 To what extent does ecPoint-rainfall post-processing correct biases in global ensemble rainfall totals, especially for convective events?
• RQ4 Does the final blended forecast product outperform all input components in terms of forecast skill and user credibility for flash flood events?
• RQ5 How does the performance of the blended system vary across different seasons and lead times, particularly for summer convective events?

Key findings

• The ecPoint-rainfall post-processed global ensemble was the most skillful of the five components (raw ECMWF, post-processed ECMWF, raw COSMO, post-processed COSMO, blended) in one-year verification.
• Post-processing of the COSMO-LAM ensemble added significant value during evening summer convective events, where the global model exhibited a systematic underprediction bias.
• In two case studies, raw COSMO overestimated the largest point rainfall totals, while raw ECMWF underestimated them; post-processing mitigated these errors in both systems.
• The final blended product appeared most credible and accurate to users, combining ecPoint’s broad-scale skill with post-processed COSMO’s localized detail.
• The scale-selective neighborhood technique successfully identified reliable small-scale features, particularly in topographically complex regions and near frontal boundaries.
• The blended forecast maintained high skill beyond 60 hours, though value added by the LAM component diminished with lead time, consistent with findings from other high-resolution LAM systems.