A low-cost post-processing technique improves weather forecasts around the world

Numerical weather prediction (NWP) models, commonly employing ensembles to represent uncertainties, currently use gridboxes spanning roughly 20km x 20km. This spatial resolution leads to a significant forecasting challenge: the inability to predict weather at specific sites, instead providing average values for much larger areas. This disconnect is especially problematic in areas with high sub-grid variability—the variation of weather conditions within a single gridbox. High-resolution models (e.g., ~2km x 2km) can reduce this issue, but their computational demands limit their geographical coverage. Calibrated post-processing (PP) techniques offer an alternative, statistically converting gridbox forecasts into point forecasts. However, existing PP methods often struggle with several challenges, including the need for extensive historical data for calibration, difficulties in representing extreme events accurately, and limited global coverage (as summarized in Table 1 of the paper). This research addresses these shortcomings by introducing a novel post-processing method.

The authors review existing methods for addressing sub-grid variability in weather forecasting. High-resolution models, while providing more realistic spatial patterns and improved forecast skill, are computationally expensive and limited in geographical scope. Existing post-processing techniques have shown some success in improving forecasts of dry weather, but face challenges in handling sub-grid variability, especially for extreme rainfall events. These challenges include the need for long (20+ years) observational and re-forecast datasets for calibration, difficulties in accurately representing the tails of probability distributions, and struggles to improve forecasts of extreme events. The authors highlight the limitations of previous approaches and position their new method, ecPoint, as a solution to these challenges.

The ecPoint method uses a non-local gridbox-analogue approach, grounded in the principles of conditional verification. It leverages the characteristics of NWP gridbox forecast outputs and other global datasets to anticipate the degree of sub-grid variability. The method categorizes gridboxes into distinct 'weather types' based on governing variables such as the convective rainfall fraction and 700 hPa wind speed. Each weather type is associated with a specific probability density function (PDF) of point rainfall within the gridbox, obtained through calibration. The calibration process uses global rain gauge observations and short-range control run re-forecasts to create a mapping function for each weather type. This non-local approach avoids the need for extensive location-specific data, allowing for global coverage and improved calibration datasets. The method also incorporates bias correction based on the weather type. For ensemble forecasts, ecPoint applies the process to each ensemble member separately, producing an 'ensemble of ensembles' which is then merged to generate the final probabilistic point forecast. The operational ecPoint system currently utilizes 214 weather types, defined in a decision tree structure. The forecast error ratio (FER) is a key metric in the calibration process, capturing the ratio of observed-to-forecast rainfall, adjusted by the forecast itself. This allows for the estimation of bias correction factors for each weather type and the generation of probabilistic point rainfall predictions.

The study demonstrates significant improvements in forecast skill using ecPoint compared to raw NWP forecasts. Verification using 1 year of retrospective forecasts and global gauge observations shows substantial gains in reliability and discrimination, particularly for extreme rainfall events (≥50 mm/12h). The improvement in reliability is particularly noticeable for the low rainfall threshold (≥0.2 mm/12h), while the large gains for high totals relate to the weather-type-dependent inclusion of multipliers for the raw NWP forecast. The ecPoint system extends the lead time with potential predictive strength by approximately 1, 2, and 8 days for 0.2 mm, 10 mm, and 50 mm thresholds respectively, representing substantial gains compared to historical improvements of about 1 day per decade. ROCA (Area under the Relative Operating Characteristic curve) scores consistently show ecPoint outperforming the raw ensemble across various lead times and rainfall thresholds. Analysis of ROC curves reveals that the added value primarily stems from better handling of the wet tails of rainfall distributions. The improvements are consistent across tropical and extratropical regions, with slightly better performance in the tropics due to the higher frequency of convective weather types. The study includes case studies illustrating the improved performance of ecPoint in predicting extreme rainfall events, such as the February 2019 flooding in Crete and a cyclonic, convective event in Norway. These case studies highlight how ecPoint produces smoother, more consistent, and geographically focused probability fields, leading to more accurate predictions of extreme events. Comparisons with other methods show that ecPoint either matches or outperforms other existing methods, including convection-resolving limited area ensembles with modern post-processing.

The results demonstrate the significant value of ecPoint as a novel post-processing technique. Its success is linked to its unique approach of using gridbox-weather-types for calibration, leading to vast and physically meaningful datasets. This differs from other methods that rely on location-specific calibration, which limits dataset size and applicability. The non-local nature of the calibration also allows for predictions in areas with limited historical data. While some regional performance variations are expected, these can be addressed by improving the governing variables and decision tree in future iterations. The method's ability to identify and correct for gridbox-scale biases provides crucial insights into model limitations, potentially aiding in model development. The superior performance of ecPoint in forecasting extreme events is especially relevant for applications such as flash flood prediction. The physically grounded nature of ecPoint contrasts with many purely statistical methods, allowing for a better interpretation of results and greater transparency. The relatively low computational cost of ecPoint makes it a practical and widely applicable tool.

The ecPoint post-processing technique offers a significant advancement in weather forecasting, particularly for point rainfall predictions. Its ability to improve reliability and skill, especially for extreme rainfall events, and its global applicability represent major contributions to the field. The approach utilizes a computationally efficient and innovative methodology that leverages global datasets for calibration, offering a scalable solution for improving weather forecasts worldwide. Future research could focus on incorporating additional governing variables, optimizing the decision tree structure using machine learning techniques, and exploring the application of ecPoint to other meteorological variables. Further integration with high-resolution models and ongoing improvements to the underlying NWP models will further enhance forecasting capabilities.

While ecPoint shows substantial improvements, some limitations exist. The reliance on rain gauge data for calibration means that the accuracy is affected by the inherent limitations of rain gauge measurements, particularly in regions with sparse gauge networks. The method's performance in regions with highly unusual topography or meteorological characteristics might be less accurate compared to location-specific methods. While the current system uses 214 weather types, further refinements to the decision tree and the inclusion of more governing variables could potentially lead to even greater accuracy, but will require increased observation coverage to support the enhanced complexity. The potential impact of outliers in training data should also be further investigated.