Numerical Weather Prediction Forecasts Research Articles

Using Machine Learning to Predict Convection-Allowing Ensemble Forecast Skill: Evaluation with the NSSL Warn-on-Forecast System

Abstract Forecasters routinely calibrate their confidence in model forecasts. Ensembles inherently estimate forecast confidence but are often underdispersive, and ensemble spread does not strongly correlate with ensemble-mean error. The misalignment between ensemble spread and skill motivates new methods for “forecasting forecast skill” so that forecasters can better utilize ensemble guidance. We have trained logistic regression and random forest models to predict the skill of composite reflectivity forecasts from the NSSL Warn-on-Forecast System (WoFS), a 3-km ensemble that generates rapidly updating forecast guidance for 0–6-h lead times. The forecast skill predictions are valid at 1-, 2-, or 3-h lead times within localized regions determined by the observed storm locations at analysis time. We use WoFS analysis and forecast output and NSSL Multi-Radar/Multi-Sensor composite reflectivity for 106 cases from the 2017 to 2021 NOAA Hazardous Weather Testbed Spring Forecasting Experiments. We frame the prediction task as a multiclassification problem, where the forecast skill labels are determined by averaging the extended fraction skill scores (eFSSs) for several reflectivity thresholds and verification neighborhoods and then converting to one of three classes based on where the average eFSS ranks within the entire dataset: POOR (bottom 20%), FAIR (middle 60%), or GOOD (top 20%). Initial machine learning (ML) models are trained on 323 predictors; reducing to 10 or 15 predictors in the final models only modestly reduces skill. The final models substantially outperform carefully developed persistence- and spread-based models and are reasonably explainable. The results suggest that ML can be a valuable tool for guiding user confidence in convection-allowing (and larger-scale) ensemble forecasts. Significance Statement Some numerical weather prediction (NWP) forecasts are more likely to verify than others. Forecasters often recognize situations where NWP output should be trusted more or less than usual, but objective methods for “forecasting forecast skill” are notably lacking for thunderstorm-scale models. Better estimates of forecast skill can benefit society through more accurate forecasts of high-impact weather. Machine learning (ML) provides a powerful framework for relating forecast skill to the characteristics of model forecasts and available observations over many previous cases. ML models can leverage these relationships to predict forecast skill for new cases in real time. We demonstrate the effectiveness of this approach to forecasting forecast skill using a cutting-edge thunderstorm prediction system and logistic regression and random forest models. Based on this success, we recommend the adoption of similar ML-based methods for other prediction models.

Impact of Assimilating Radar Solid Precipitation Data in the Canadian Precipitation Analysis System

Abstract The Canadian Precipitation Analysis (CaPA) system provides near-real-time precipitation analyses over Canada by combining observations with short-term numerical weather prediction forecasts. CaPA’s snowfall estimates suffer from the lack of accurate solid precipitation measurements to correct the first-guess estimate. Weather radars have the potential to add precipitation measurements to CaPA in all seasons but are not assimilated in winter due to radar snowfall estimate imprecision and lack of precipitation gauges for calibration. The main objective of this study is to assess the impact of assimilating Canadian dual-polarized radar-based snowfall data in CaPA to improve precipitation estimates. Two sets of experiments were conducted to evaluate the impact of including radar snowfall retrievals, one set using the high-resolution CaPA (HRDPA) with the currently operational quality control configuration and another increasing the number of assimilated surface observations by relaxing quality control. Experiments spanned two winter seasons (2021 and 2022) in central Canada, covering part of the entire CaPA domain. The results showed that the assimilation of radar-based snowfall data improved CaPA’s precipitation estimates 81.75% of the time for 0.5-mm precipitation thresholds. An increase in the probability of detection together with a decrease in the false alarm ratio suggested an improvement of the precipitation spatial distribution and estimation accuracy. Additionally, the results showed improvements for both precipitation mass and frequency biases for low precipitation amounts. For larger thresholds, the frequency bias was degraded. The results also indicated that the assimilation of dual-polarization radar data is beneficial for the two CaPA configurations tested in this study.

Spatial solar forecast verification with the neighborhood method and automatic threshold segmentation

Numerical weather prediction (NWP) has hitherto been the default tool for providing day-ahead forecasting services to the solar energy industry. Rapid advancements in solar forecasting using NWP call for more appropriate forecast verification procedures. Current solar forecast verification is almost always carried out through ground-based radiometric data collected at point locations. Consequently, spatial features embedded in the gridded NWP forecasts cannot be verified. This study presents the spatial verification of solar irradiance forecasts using the neighborhood method, with the main goal of emphasizing the importance of such verification procedures. By applying spatial smoothing one establishes a way to directly compare the observed and forecast fields, and concurrently, mitigate verification errors that may arise from small-scale spatial displacements. Within this framework, two variants of the neighborhood-based verification, namely, the fraction-field method and the upscaling method, are examined with respect to two reanalysis products, namely, ERA5 and MERRA-2. The results suggest that, in comparison to the upscaling method, the fraction-field method can better quantify forecast performance by providing fractions skill scores. On top of the traditional neighborhood approach, which involves the subjective selection of threshold for dichotomization, an automatic threshold segmentation method based on the three-component skew-normal mixture model is proposed to resolve the issue, which can also lead to substantial time savings in data processing. Given the spatio-temporal attributes and benefits of visualization, spatial verification is anticipated to serve as a complementary practice to the current mainstream point-location forecast verification.

Assimilation of Transformed Retrievals from IASI radiances versus direct assimilation of IASI radiances at the Met Office

AbstractSatellite radiances make up the vast majority of observations that are assimilated routinely at the Met Office and at other major numerical weather prediction (NWP) centres. Improvements in the use of these data have led to significant improvements in NWP forecast skill over the last 30 years. Since the advent of hyperspectral sensors, capable of measuring infrared radiance emerging from the atmosphere over thousands of channels, data providers and NWP centres have been facing the challenges of distributing, monitoring, and assimilating a considerable wealth of data. Forthcoming hyperspectral sensors such as the Infrared Atmospheric Sounding Interferometer New Generation (IASI‐NG) on MetOp Second Generation (Metop‐SG) and the Meteosat Third Generation Infrared Sounder (MTG‐IRS) will increase the amount of data to be considered for assimilation even further. It has been shown that Transformed Retrievals (TRs) generated from a spectrum of radiances can represent most of the information content of the radiances in a significantly reduced set of TR components. In this article we show for the first time, to our knowledge, that by assimilating TRs in an operational data assimilation system it is possible to achieve forecast skill gains similar to those when the same radiances used to generate the TRs are assimilated directly. These results pave the way to exploiting an order‐of‐magnitude larger number of channels from hyperspectral sounders for operational NWP.

Near-Surface Thermodynamic Influences on Evaporation Duct Shape

This study utilizes in situ measurements and numerical weather prediction forecasts curated during the Coupled Air–Sea Processes Electromagnetic Ducting Research (CASPER) east field campaign to assess how thermodynamic properties in the marine atmospheric surface layer influence evaporation duct shape independent of duct height. More specifically, we investigate evaporation duct shape through a duct shape parameter, a parameter known to affect the propagation of X-band radar signals and is directly related to the curvature of the duct. Relationships between this duct shape parameter and air sea temperature difference (ASTD) reveal that during unstable periods (ASTD < 0), the duct shape parameter is generally larger than in near-neutral or stable atmospheric conditions, indicating tighter curvature of the M-profile. Furthermore, for any specific duct height, a strong linear relationship between the near-surface-specific humidity gradient and the duct shape parameter is found, suggesting that it is primarily driven by near-surface humidity gradients. The results demonstrate that an a priori estimate of duct shape, for a given duct height, is possible if the near-surface humidity gradient is known.

Bias correction of wind power forecasts with SCADA data and continuous learning

Wind energy plays a critical role in the transition towards renewable energy sources. However, the uncertainty and variability of wind can impede its full potential and the necessary growth of wind power capacity. To mitigate these challenges, wind power forecasting methods are employed for applications in power management, electricity trading, or maintenance scheduling. In this work, we present, evaluate, and compare four machine learning-based wind power forecasting models. Our models correct and improve 48-hour forecasts extracted from a numerical weather prediction (NWP) model. The models are evaluated on datasets from a wind park comprising 65 wind turbines. The best improvement in forecasting error and mean bias was achieved by a convolutional neural network, reducing the average NRMSE down to 22%, coupled with a significant reduction in mean bias, compared to a NRMSE of 35% from the strongly biased baseline model using uncorrected NWP forecasts. Our findings further indicate that changes to neural network architectures play a minor role in affecting the forecasting performance, and that future research should rather investigate changes in the model pipeline. Moreover, we introduce a continuous learning strategy, which is shown to achieve the highest forecasting performance improvements when new data is made available.

Downscaled numerical weather predictions can improve forecasts of sugarcane irrigation indices

Efficient irrigation reduces energy and water costs, increases profit margins and delivers better environmental outcomes. Whilst many growers rely on weather forecasts to make decisions, few studies have sought to incorporate weather forecast uncertainty into the optimisation of irrigation management or to evaluate weather forecasts through the lens of irrigation indices. Therefore, in this study, we seek to generate and evaluate ensemble forecasts of irrigation indices produced by coupling numerical weather prediction (NWP) forecasts with a biophysical process model (APSIM). We investigate a case study application for sugarcane in northeastern Australia. As a first step, three and a half years of forecasts from the Australian Bureau of Meteorology’s ACCESS-G3 model are statistically post-processed to generate 7-day forecasts that are downscaled and calibrated to local climate zones. In addition, the forecast post-processor converts the deterministic forecasts into an ensemble, thus quantifying forecast uncertainty. The generated forecasts are then used as forcing for the APSIM crop model to produce ensemble forecasts of soil water deficit (SWD), crop water use (CWU) and crop stress (Stress) for a simulated sugarcane crop. Through cross-validation, the post-processed weather forecasts demonstrate improved forecast accuracy compared to naive climatology and raw NWP forecasts for daily rainfall, maximum and minimum temperature and solar radiation; with the added benefit of providing a reliable uncertainty estimate. Improvement to an even greater degree is observed for the derived irrigation indices, particularly CWU and Stress, for which the forecasts are also reliable. The developed irrigation indices based on NWP can be used directly for decision making or, alternatively, may be used further in machine learning for optimisation of irrigation schedules in conjunction with other remotely sensed variables.

NYSolarCast: A solar power forecasting system for New York State

Solar energy generation capacity will need to be greatly increased to meet aggressive clean energy targets by New York State (NYS), which are 70% renewable energy (RE) generation by 2030, and 100% by 2040. Because solar energy is a variable, weather-driven resource, accurate forecasts of solar energy generation both in nowcast (intra-day) and day-ahead time horizons are necessary for electric utilities and independent system operators, as they maintain grid stability and maximize RE use.Meeting this need for NYS, a gridded, open-source solar power forecasting system called NYSolarCast was developed (https://github.com/ncar/NYSolarCast_delivery). NYSolarCast makes 15-min resolution predictions of global horizontal irradiance (GHI) on a 3-km grid covering all of NYS, which are then used to estimate solar power generation both for select utility-scale PV plants (15-min resolution) and for zone-aggregated distributed PV (1-h resolution). The statewide GHI forecasts are made by applying the StatCast statistical forecasting model, which is trained on over two years of GHI observations from pyranometers at all 126 NYS Mesonet stations and gridded numerical weather prediction forecasts from both the Weather Research and Forecasting model tuned for solar applications (WRF-Solar®) and NOAA’s operational High-Resolution Rapid Refresh (HRRR) model. Forecasts are made at each NYS Mesonet site and then blended outward into the rest of the grid. This paper gives an overview of NYSolarCast performance for intra-day (0–6-h) GHI and power forecasts during a one-year period.

Bayesian physical–statistical retrieval of snow water equivalent and snow depth from X- and Ku-band synthetic aperture radar – demonstration using airborne SnowSAr in SnowEx'17

Abstract. A physical–statistical framework to estimate snow water equivalent (SWE) and snow depth from synthetic aperture radar (SAR) measurements is presented and applied to four SnowSAR flight-line data sets collected during the SnowEx'2017 field campaign in Grand Mesa, Colorado, USA. The physical (radar) model is used to describe the relationship between snowpack conditions and volume backscatter. The statistical model is a Bayesian inference model that seeks to estimate the joint probability distribution of volume backscatter measurements, snow density and snow depth, and physical model parameters. Prior distributions are derived from multilayer snow hydrology predictions driven by downscaled numerical weather prediction (NWP) forecasts. To reduce the signal-to-noise ratio, SnowSAR measurements at 1 m resolution were upscaled by simple averaging to 30 and 90 m resolution. To reduce the number of physical parameters, the multilayer snowpack is transformed for Bayesian inference into an equivalent one- or two-layer snowpack with the same snow mass and volume backscatter. Successful retrievals meeting NASEM (2018) science requirements are defined by absolute convergence backscatter errors ≤1.2 dB and local SnowSAR incidence angles between 30 and 45∘ for X- and Ku-band VV-pol backscatter measurements and were achieved for 75 % to 87 % of all grassland pixels with SWE up to 0.7 m and snow depth up to 2 m. SWE retrievals compare well with snow pit observations, showing strong skill in deep snow with average absolute SWE residuals of 5 %–7 % (15 %–18 %) for the two-layer (one-layer) retrieval algorithm. Furthermore, the spatial distributions of snow depth retrievals vis-à-vis lidar estimates have Bhattacharya coefficients above 94 % (90 %) for homogeneous grassland pixels at 30 m (90 m resolution), and values up to 76 % in mixed forest and grassland areas, indicating that the retrievals closely capture snowpack spatial variability. Because NWP forecasts are available everywhere, the proposed approach could be applied to SWE and snow depth retrievals from a dedicated global snow mission.

ПІДВИЩЕННЯ ТОЧНОСТІ БАГАТОФАКТОРНИХ КОРОТКОСТРОКОВИХ ПРОГНОЗІВ ГЕНЕРАЦІЇ СЕС НА ОСНОВІ ШТУЧНОЇ НЕЙРОННОЇ МЕРЕЖІ

The paper focuses on the development of models for forecasting the electricity generation of industrial solar power plants using artificial neural networks and numerical weather prediction. The relevance of the research is driven by the need to reduce costs related to imbalances in electricity generation from renewable sources, which can sometimes reach 50% of the released electricity. Additionally, the imbalances of such producers are increasing in Ukraine's power system. Currently, the general imbalances of renewable energy producers in Ukraine have led to a 45% reduction in green electricity production, especially due to the damage or destruction of 75% of wind power plants and 15% of solar power plants in southern and southeastern regions as a result of hostilities. Increasing the accuracy and stability of electricity generation forecasts for such producers could significantly reduce costs associated with imbalances.. Various aggregation methods have been developed for 15-minute values of green energy generation to enhance forecasting accuracy for 1, 2, and 24-hour intervals. The study investigated the potential benefits of using numerical weather prediction (NWP) forecast values to enhance forecasting accuracy. The study revealed the significance of different factors for forecasting at each bias interval. The study employed two modern recurrent neural network models, LSTM and GRU, with varying time sequences. References 14, figures 5, table 2.

Nowcasting convective activity for the Sahel: A simple probabilistic approach using real‐time and historical satellite data on cloud‐top temperature

AbstractFlash flooding from intense rainfall frequently results in major damage and loss of life across Africa. In the Sahel, automatic prediction and warning systems for these events, driven by Mesoscale Convective Systems (MCSs), are limited, and Numerical Weather Prediction (NWP) forecasts continue to have little skill. The ground observation network is also sparse, and very few operational meteorological radars exist to facilitate conventional nowcasting approaches. Focusing on the western Sahel, we present a novel approach for producing probabilistic nowcasts of convective activity out to six hours ahead, using the current location of observed convection. Convective parts of the MCS, associated with extreme and heavy precipitation, are identified from 16 years of Meteosat Second Generation thermal–infrared cloud‐top temperature data, and an offline database of location‐conditioned probabilities calculated. From this database, real‐time nowcasts can be quickly produced with minimal calculation. The nowcasts give the probability of convection occurring within a square neighbourhood surrounding each grid point, accounting for the inherent unpredictability of convection at small scales. Compared to a climatological reference, formal verification approaches show the nowcasts to be skilful at predicting convective activity over the study region, for all times of day and out to the six‐hour lead time considered. The nowcasts are also skilful at capturing extreme 24‐hour rain gauge accumulations over Dakar, Senegal. The nowcast skill peaks in the afternoon, with a minimum in the evening. We find that the optimum neighbourhood size varies with lead time, from 10 km at the nowcast origin to around 100 km at a six‐hour lead time. This simple and skilful nowcasting method could be highly valuable for operational warnings across West Africa and other regions with long‐lived thunderstorms, and help to reduce the impacts from heavy rainfall and flooding.

Cold Fog Amongst Complex Terrain

Abstract Cold fog forms via various thermodynamic, dynamic, and microphysical processes when the air temperature is less than 0°C. It occurs frequently during the cold season in the western United States yet is challenging to detect using standard observations and is very difficult to predict. The Cold Fog Amongst Complex Terrain (CFACT) project was conceived to investigate the life cycle of cold fog in mountain valleys. The overarching goals of the CFACT project are to 1) investigate the life cycle of cold-fog events over complex terrain with the latest observation technology, 2) improve microphysical parameterizations and visibility algorithms used in numerical weather prediction (NWP) models, and 3) develop data assimilation and analysis methods for current and next-generation (e.g., subkilometer scale) NWP models. The CFACT field campaign took place in Heber Valley, Utah, during January and February 2022, with support from NSF’s Lower Atmospheric Observing Facilities (managed by NCAR’s Earth Observing Laboratory), the University of Utah, and Ontario Technical University. A network of ground-based and aerial in situ instruments and remote sensing platforms were used to obtain comprehensive measurements of thermodynamic profiles, cloud microphysics, aerosol properties, and environmental dynamics. Nine intensive observation periods (IOPs) explored various mountainous weather and cold-fog conditions. Field observations, NWP forecasts, and large-eddy simulations provided unprecedented data sources to help understand the mechanisms associated with cold-fog weather and to identify and mitigate numerical model deficiencies in simulating winter weather over mountainous terrain. This article summarizes the CFACT field campaign, its observations, and challenges during the field campaign, including real-time fog prediction issues and future analysis.

Investigating transformer‐based models for spatial downscaling and correcting biases of near‐surface temperature and wind‐speed forecasts

AbstractHigh‐resolution and accurate prediction of near‐surface weather parameters based on numerical weather prediction (NWP) models is essential for many downstream and real‐world applications. Traditional dynamical or statistical downscaling methods are insufficient to derive high‐resolution data from operational NWP forecasts, making it essential to devise new approaches. In recent years, an increasing number of researchers have explored the implementations of deep learning (DL) based models for spatial downscaling, motivated by the similarity between the super‐resolution (SR) problem in computer vision (CV) and downscaling. Furthermore, while transformer‐based models have become state‐of‐the‐art models for many SR tasks, they are rarely applied for downscaling of weather forecasts or climate projections. This study adapted transformer‐based models such as SwinIR and Uformer to downscale the temperature at 2 m () and wind speed at 10 m () over Eastern Inner Mongolia, encompassing the area from 39.6–46°N latitude and 111.6–118°E longitude. We used high‐resolution forecast (HRES) data from the European Centre for Medium‐range Weather Forecast (ECMWF) with a spatial resolution of 0.1° as the input and gridded observation data from the China Meteorological Administration (CMA) Land Data Assimilation System (CLDAS) at a spatial resolution of 0.01° as the target. Given that the models use observation data rather than a coarse‐grained version of forecast data as the target, they accomplish both bias correction and spatial downscaling. The results demonstrate that the performance of SwinIR and Uformer is superior to that of two convolutional neural network (CNN) based models (UNet and RCAN). Additionally, we introduced a novel module to extract features of varying resolution from the high‐resolution topography data and applied a multiscale feature fusion module to merge features of different scales, contributing to further enhancement of Uformer's performance.

Using synthetic case studies to explore the spread and calibration of ensemble atmospheric dispersion forecasts

Abstract. Ensemble predictions of atmospheric dispersion that account for the meteorological uncertainties in a weather forecast are constructed by propagating the individual members of an ensemble numerical weather prediction forecast through an atmospheric dispersion model. Two event scenarios involving hypothetical atmospheric releases are considered: a near-surface radiological release from a nuclear power plant accident and a large eruption of an Icelandic volcano releasing volcanic ash into the upper air. Simulations were run twice-daily in real time over a 4-month period to create a large dataset of cases for this study. The performance of the ensemble predictions is measured against retrospective simulations using a sequence of meteorological fields analysed against observations. The focus of this paper is on comparing the spread of the ensemble members against forecast errors and on the calibration of probabilistic forecasts derived from the ensemble distribution. Results show good overall performance by the dispersion ensembles in both studies but with simulations for the upper-air ash release generally performing better than those for the near-surface release of radiological material. The near-surface results demonstrate a sensitivity to the release location, with good performance in areas dominated by the synoptic-scale meteorology and generally poorer performance at some other sites where, we speculate, the global-scale meteorological ensemble used in this study has difficulty in adequately capturing the uncertainty from local- and regional-scale influences on the boundary layer. The ensemble tends to be under-spread, or over-confident, for the radiological case in general, especially at earlier forecast steps. The limited ensemble size of 18 members may also affect its ability to fully resolve peak values or adequately sample outlier regions. Probability forecasts of threshold exceedances show a reasonable degree of calibration, though the over-confident nature of the ensemble means that it tends to be too keen on using the extreme forecast probabilities. Ensemble forecasts for the volcanic ash study demonstrate an appropriate degree of spread and are generally well-calibrated, particularly for ash concentration forecasts in the troposphere. The ensemble is slightly over-spread, or under-confident, within the troposphere at the first output time step T + 6, thought to be attributable to a known deficiency in the ensemble perturbation scheme in use at the time of this study, but improves with probability forecasts becoming well-calibrated here by the end of the period. Conversely, an increasing tendency towards over-confident forecasts is seen in the stratosphere, which again mirrors an expectation for ensemble spread to fall away at higher altitudes in the meteorological ensemble. Results in the volcanic ash case are also broadly similar between the three different eruption scenarios considered in the study, suggesting that good ensemble performance might apply to a wide range of eruptions with different heights and mass eruption rates.

Deep-learning post-processing of short-term station precipitation based on NWP forecasts

Post-processing methods that rely on fusion-grided forecast products can reduce systematic biases from Numerical Weather Prediction (NWP) precipitation forecasts. However, these methods also limit the capability to forecast precipitation accurately at local stations. We constructed a Station-based Precipitation Post-processing Model (SPPM) that utilizes deep-learning algorithms, predominantly convolutional layers and ResNet modules. Based on 390 meteorological stations in North China and European Centre for Medium-Range Weather Forecasts Highest-resolution (ECMWF-HRES) forecast data, the SPPM utilizes multi-level atmospheric forecast variables and geographic variables in a small area centered on a station as predictors. The results show that the SPPM improved the threat score (TS) by 4.29%, 3.66%, 15.63%, 61.08%, and 295.83% for precipitation thresholds of 0.1, 3.0, 10.0, 20.0, and 50.0 mm/3 h, respectively. We then examined the sensitivity of predictors using the interpretable deep-learning technique Layer-wise Relevance Propagation (LRP). The results indicate that the NWP total precipitation (TP) from ECMWF is the most sensitive and important factor, followed by the low-level (850 hPa) field, single-level field, and geographic variables. Notably, TP becomes increasingly important with larger forecast grades, while the importance of variables at other levels remains relatively constant. The majority of stations exhibit consistent importance rankings as mentioned above. Finally, possible causes of variables' insensitivity at medium-level (500 hPa) and high-level (200 hPa) were discussed.

A New Method for Postprocessing Numerical Weather Predictions Using Quantile Mapping in the Frequency Domain

Abstract Improving lead time for forecasting floods is important to minimize property damage and ensure the safety of the public and emergency services during flood events. Numerical weather prediction (NWP) models are important components of flood forecasting systems and have been vital in extending forecasting lead time under complex weather and terrain conditions. However, NWP forecasts still have significant uncertainty associated with the precipitation fields that are the main inputs of the hydrologic models and thus the resulting flood forecasts. An issue often overlooked is the importance of correctly representing variability over a range of different temporal scales. To address this gap, here a new wavelet-based method for postprocessing NWP precipitation forecasts is proposed. First, precipitation forecasts are decomposed into the frequency domain using a wavelet transform, providing estimates of the amplitudes and phases of the time series at different frequencies. Quantile mapping is then used to correct bias in the amplitudes of each frequency. Randomized phases are used to generate an ensemble of realizations of the precipitation forecasts. The postprocessed precipitation forecasts are reconstructed by taking the inverse of adjusted time-frequency decompositions with the corrected amplitudes and randomized phases. The proposed method was used to postprocess NWP precipitation forecasts in the Sydney region of Australia. There is a significant improvement in postprocessed precipitation forecasts across multiple time scales in terms of bias and temporal and spatial correlation structures. The postprocessed precipitation fields can be used for the modeling of fully distributed hydrologic systems, improving runoff stimulation, flood depth estimation, and flood early warning. Significance Statement A new method accounting for the timing and spatial errors of NWP precipitation forecasts is proposed, and it can improve the skill of forecasts across multiple time scales, especially at short lead times. The proposed method provides a practical and effective way to correct these errors by incorporating spatiotemporal neighborhood information through the frequency domain using sophisticated wavelet transforms. With systematic timing and spatial errors removed, precipitation forecasts will be more skillful, and hydrological modeling using the postprocessed forecasts can provide higher accuracy of streamflow estimation.

Assessments of Use of Blended Radar–Numerical Weather Prediction Product in Short-Range Warning of Intense Rainstorms in Localized Systems (SWIRLS) for Quantitative Precipitation Forecast of Tropical Cyclone Landfall on Vietnam’s Coast

This research presents a blended system implemented by the Vietnam National Center for Hydro-Meteorological Forecasting to enhance the nowcasting and forecasting services of quantitative precipitation forecasts (QPFs) of tropical cyclone (TC) landfalls on Vietnam’s coast. Firstly, the extrapolations of rain/convective systems from multiple radars in Vietnam in ranges up to 6 h were carried out using Short-Range Warning of Intense Rainstorms in Localized Systems (SWIRLS) developed by the Hong Kong Observatory. Secondly, the forecast from the numerical weather prediction (NWP) system, based on the WRF-ARW model running at 3 km horizontal resolution, was blended with radar-based quantitative precipitation estimates and nowcasts of SWIRLS. The analysis showed that the application of the nowcast system to TC-related cloud forms is complicated, which is related to the TC’s evolution and the different types and multiple layers of storm clouds that can affect the accuracy of the derived motion fields in nowcast systems. With hourly accumulated rainfall observation, skill score validation conducted for several TCs that landed in the center of Vietnam demonstrated that the blending of nowcasting and NWP improve the quality of the QPFs of TCs in forecast ranges up to 3 h compared to the pure NWP forecasts.

Correcting Subseasonal Forecast Errors with an Explainable ANN to Understand Misrepresented Sources of Predictability of European Summer Temperatures

Abstract Subseasonal forecasts are challenging for numerical weather prediction (NWP) and machine learning models alike. Forecasting 2-m temperature (t2m) with a lead time of 2 or more weeks requires a forward model to integrate multiple complex interactions, like oceanic and land surface conditions leading to predictable weather patterns. NWP models represent these interactions imperfectly, meaning that in certain conditions, errors accumulate and model predictability deviates from real predictability, often for poorly understood reasons. To advance that understanding, this paper corrects conditional errors in NWP forecasts with an artificial neural network (ANN). The ANN postprocesses ECMWF extended-range summer temperature forecasts by learning to correct the ECMWF-predicted probability that monthly t2m in western and central Europe exceeds the climatological median. Predictors are objectively selected from ECMWF forecasts themselves, and from states at initialization, i.e., the ERA5 reanalysis. The latter allows the ANN to account for sources of predictability that are biased in the NWP model itself. We attribute ANN corrections with two explainable artificial intelligence (AI) tools. This reveals that certain erroneous forecasts relate to tropical western Pacific Ocean sea surface temperatures at initialization. We conjecture that the atmospheric teleconnection following this source of predictability is imperfectly represented by the ECMWF model. Correcting the associated conditional errors with the ANN improves forecast skill. Significance Statement We want to understand occasions in which a numerical weather prediction (NWP) model fails to forecast a predictable event existing in the real world. For forecasts of European summer weather more than 2 weeks in advance, real predictable events are rare. When misrepresented by the model, predicted future states become needlessly biased. We diagnose these missed opportunities with an explainable neural network. The neural network is aware of the initial state and learns to correct the NWP forecast on occasions when it misrepresents a teleconnection from the western tropical Pacific Ocean to Europe. The explainable architecture can be useful for other applications in which conditional model errors need to be understood and corrected.

GEFSv12 High- and Low-Skill Day-10 Tornado Forecasts

Abstract On average, modern numerical weather prediction forecasts for daily tornado frequency exhibit no skill beyond day 10. However, in this extended-range lead window, there are particular model cycles that have exceptionally high forecast skill for tornadoes because of their ability to correctly simulate the future synoptic pattern. Here, model initial conditions that produced a more skillful forecast for tornadoes over the United States were exploited while also highlighting potential causes for low-skill cycles within the Global Ensemble Forecasting System, version 12 (GEFSv12). There were 88 high-skill and 91 low-skill forecasts in which the verifying day-10 synoptic pattern for tornado conditions revealed a western U.S. thermal trough and an eastern U.S. thermal ridge, a favorable configuration for tornadic storm occurrence. Initial conditions for high skill forecasts tended to exhibit warmer sea surface temperatures throughout the tropical Pacific Ocean and Gulf of Mexico, an active Madden–Julian oscillation, and significant modulation of Earth-relative atmospheric angular momentum. Low-skill forecasts were often initialized during La Niña and negative Pacific decadal oscillation conditions. Significant atmospheric blocking over eastern Russia—in which the GEFSv12 overforecast the duration and characteristics of the downstream flow—was a common physical process associated with low-skill forecasts. This work helps to increase our understanding of the common causes of high- or low-skill extended-range tornado forecasts and could serve as a helpful tool for operational forecasters. Significance Statement This research provides a framework for the anticipation of a more (or less) skillful 10-day tornado forecast in an operational numerical weather prediction system. High-skill forecasts were associated with substantial tropical convection and warm sea surface temperature throughout the Pacific Ocean and Gulf of Mexico, whereas the underlying cause of low-skill forecasts were typically associated with a blocking anticyclone over eastern Russia. These findings are important because they permit increased or decreased confidence in a long-range forecast of tornado occurrence based on a dynamical prediction system.

Scale‐dependent blending of ensemble rainfall nowcasts and numerical weather prediction in the open‐source pysteps library

AbstractFlash flood early warning requires accurate rainfall forecasts with a high spatial and temporal resolution. As the first few hours ahead are already not sufficiently well captured by the rainfall forecasts of numerical weather prediction (NWP) models, radar rainfall nowcasting can provide an alternative. Because this observation‐based method quickly loses skill after the first 2 hr of the forecast, it needs to be combined with NWP forecasts to extend the skillful lead time of short‐term rainfall forecasts, which should increase decision‐making times. We implemented an adaptive scale‐dependent ensemble blending method in the open‐source pysteps library, based on the Short‐Term Ensemble Prediction System scheme. In this implementation, the extrapolation (ensemble) nowcast, (ensemble) NWP, and noise components are combined with skill‐dependent weights that vary per spatial scale level. To constrain the (dis)appearance of rain in the ensemble members to regions around the rainy areas, we have developed a Lagrangian blended probability matching scheme and incremental masking strategy. We describe the implementation details and evaluate the method using three heavy and extreme (July 2021) rainfall events in four Belgian and Dutch catchments. We benchmark the results of the 48‐member blended forecasts against the Belgian NWP forecast, a 48‐member nowcast, and a simple 48‐member linear blending approach. Both on the radar domain and catchment scale, the introduced blending approach predominantly performs similarly or better than only nowcasting (in terms of event‐averaged continuous ranked probability score and critical success index values) and adds value compared with NWP for the first hours of the forecast, although the difference, particularly with the linear blending method, reduces when we focus on catchment‐average cumulative rainfall sums instead of instantaneous rainfall rates. By properly combining observations and NWP forecasts, blending methods such as these are a crucial component of seamless prediction systems.