A Comparative Analysis of Data-Driven Machine Learning Models for Rainfall Forecasting in Bangladesh

This research evaluates the quality of high-impact rainfall forecasts across Vietnam and its sub-climate regions. The 3-day rainfall forecast products evaluated include the European Centre for Medium-Range Weather Forecasts (ECMWF) High-Resolution Integrated Forecasting System (IFS) and its downscaled outputs using the Weather Research and Forecasting (WRF) model with the Advanced Research WRF core (WRF-ARW): direct downscaling and downscaling with data assimilation. A full 5-year validation period from 2019 to 2025 was processed. The validation focused on basic rainfall thresholds and also considered the distribution of skill scores for intense events and extreme events. The validations revealed systematic errors (bias) in the models at low rainfall thresholds. The forecast skill was the lowest for northern regions, while the central regions exhibited the highest. For regions strongly affected by terrain, high-resolution downscaling with local observation data assimilation is necessary to improve the detectability of extreme events.

This paper compares upper-tropospheric humidity from the Microwave Limb Sounder (MLS) on the Upper Atmosphere Research Satellite with European Centre for Medium-Range Weather Forecasts (ECMWF) data. MLS measurements are not included in the ECMWF analyses, and so a comparison of two independent datasets is possible. The focus of this paper is on the tropical region from 1991 to 1997 when MLS measurements are available and, to be more specific, from 1991 to 1994 during the ECMWF 15-yr reanalysis period (ERA-15). It is found that the contrast between moist and dry areas, or areas of convection and subsidence, is less pronounced in ECMWF data than in MLS measurements. This result applies on the scale of both global and regional circulations and results in a semiannual cycle of high and low correlations between the two datasets that is particularly pronounced over South America. Time series show that the impact of model changes to the absolute values of humidity make long-term comparison with MLS difficult.

Abstract. Global seasonal forecasts of meteorological drought using the standardized precipitation index (SPI) are produced using two data sets as initial conditions: the Global Precipitation Climatology Centre (GPCC) and the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA-Interim reanalysis (ERAI); and two seasonal forecasts of precipitation, the most recent ECMWF seasonal forecast system and climatologically based ensemble forecasts. The forecast evaluation focuses on the periods where precipitation deficits are likely to have higher drought impacts, and the results were summarized over different regions in the world. The verification of the forecasts with lead time indicated that generally for all regions the least reduction on skill was found for (i) long lead times using ERAI or GPCC for monitoring and (ii) short lead times using ECMWF or climatological seasonal forecasts. The memory effect of initial conditions was found to be 1 month of lead time for the SPI-3, 4 months for the SPI-6 and 6 (or more) months for the SPI-12. Results show that dynamical forecasts of precipitation provide added value with skills at least equal to and often above that of climatological forecasts. Furthermore, it is very difficult to improve on the use of climatological forecasts for long lead times. Our results also support recent questions of whether seasonal forecasting of global drought onset was essentially a stochastic forecasting problem. Results are presented regionally and globally, and our results point to several regions in the world where drought onset forecasting is feasible and skilful.

Surface water and energy budgets from the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) and European Centre for Medium-Range Weather Forecasts (ECMWF) reanalyses are compared here with each other and with available observations over the Mississippi River basin, which is a focus of the Global Energy and Water Cycle Experiment Continental-Scale International Project. There are a number of noticeable differences and similarities in the large-scale basin averages. The NCEP–NCAR reanalysis seasonal precipitation and runoff are larger than the available observations; presumably, evaporation and surface water variations also are too large. The ECMWF reanalysis precipitation is much closer to the observations, whereas the corresponding surface runoff and seasonal surface water variations are too small. The NCEP–NCAR and ECMWF reanalysis seasonal energy components are more similar to each other. The NCEP–NCAR and ECMWF interannual variations also are comparable, indicating that these reanalyses probably can be used to begin to study interannual variations. Nonetheless, improved land surface parameterizations are needed to depict surface water and energy processes and, in particular, variations in seasonal surface water and runoff better.

Eleven Tropical Cyclone (TC) intensity guidance models in the western North Pacific have been validated over 2008~2014 based on various analysis methods according to the lead time of forecast, year, month, intensity, rapid intensity change, track, and geographical area with an additional focus on TCs that influenced the Korean peninsula. From the evaluation using mean absolute error and correlation coefficients for maximum wind speed forecasts up to 72 h, we found that the Hurricane Weather Research and Forecasting model (HWRF) outperforms all others overall although the Global Forecast System (GFS), the Typhoon Ensemble Prediction System of Japan Meteorological Agency (TEPS), and the Korean version of Weather and Weather Research and Forecasting model (KWRF) also shows a good performance in some lead times of forecast. In particular, HWRF shows the highest performance in predicting the intensity of strong TCs above Category 3, which may be attributed to its highest spatial resolution (~3 km). The Navy Operational Global Prediction Model (NOGAPS) and GFS were the most improved model during 2008~2014. For initial intensity error, two Japanese models, Japan Meteorological Agency Global Spectral Model (JGSM) and TEPS, had the smallest error. In track forecast, the European Centre for Medium-Range Weather Forecasts (ECMWF) and recent GFS model outperformed others. The present results has significant implications for providing basic information for operational forecasters as well as developing ensemble or consensus prediction systems.

This paper evaluates the extratropical cyclones within three operational global ensembles [the 20-member Canadian Meteorological Centre (CMC), 20-member National Centers for Environmental Prediction (NCEP), and 50-member European Centre for Medium-Range Weather Forecasts (ECMWF)]. The day-0–6 forecasts were evaluated over the eastern United States and western Atlantic for the 2007–15 cool seasons (October–March) using the ECMWF’s ERA-Interim dataset as the verifying analysis. The Hodges cyclone-tracking scheme was used to track cyclones using 6-h mean sea level pressure (MSLP) data. For lead times less than 72 h, the NCEP and ECMWF ensembles have comparable mean absolute errors in cyclone intensity and track, while the CMC errors are larger. For days 4–6 ECMWF has 12–18 and 24–30 h more accuracy for cyclone intensity than NCEP and CMC, respectively. All ensembles underpredict relatively deep cyclones in the medium range, with one area near the Gulf Stream. CMC, NCEP, and ECMWF all have a slow along-track bias that is significant from 24 to 90 h, and they have a left-of-track bias from 120 to 144 h. ECMWF has greater probabilistic skill for intensity and track than CMC and NCEP, while the 90-member multimodel ensemble (NCEP + CMC + ECMWF) has more probabilistic skill than any single ensemble. During the medium range, the ECMWF + NCEP + CMC multimodel ensemble has the fewest cases (1.9%, 1.8%, and 1.0%) outside the envelope compared to ECMWF (5.6%, 5.2%, and 4.1%) and NCEP (13.7%, 10.6%, and 11.0%) for cyclone intensity and along- and cross-track positions.

In this paper, the results of an examination of precipitation forecasts for 1–30-day leads from global models run at the European Centre for Medium-Range Weather Forecasts (ECMWF) and the National Centers for Environmental Prediction (NCEP) during November 2007–February 2008 are presented. The performance of the model precipitation forecasts are examined in global and regional contexts, and results of a case study of precipitation variations that are associated with a moderate to strong Madden–Julian oscillation (MJO) event are presented.The precipitation forecasts from the ECMWF and NCEP operational prediction models have nearly identical temporal correlation with observed precipitation at forecast leads from 2 to 9 days over the Northern Hemisphere during the cool season, despite the higher resolution of the ECMWF operational model, while the ECMWF operational model forecasts are slightly better in the tropics and the Southern Hemisphere during the warm season. The ECMWF Re-Analysis Interim (ERA-Interim) precipitation forecasts perform only slightly worse than the NCEP operational model, while NCEP’s Climate Forecast System low-resolution coupled model forecasts perform the worst among the four models. In terms of bias, the ECMWF operational model performs the best among the four model forecasts that were examined, particularly with respect to the ITCZ regions in both the Atlantic and Pacific. Local temporal correlations that were computed on daily precipitation totals for day-2 forecasts against observations indicate that the operational models at ECMWF and NCEP perform the best during the 4-month study period, and that all of the models have low to insignificant correlations over land and over much of the tropics. They perform the best in subtropical and extratropical oceanic regions.Also presented are results that show that striking improvements have been made over the past two decades in the ability of the models to represent precipitation variations that are associated with MJO. The model precipitation forecasts exhibit the ability to characterize the evolution of precipitation variations during a moderate–strong period of MJO conditions for forecast leads as long as 10 days.

Recent studies have shown the importance of high-resolution wind in coastal ocean modeling. This paper tests the Coupled Ocean–Atmosphere Mesoscale Prediction System (COAMPS) at the 9-, 27-, and 81-km grid resolutions in simulating wind off the central and southern California coasts, including the Santa Barbara Channel (SBC). The test period is March–May (1999) when the wind changes from its characteristics more typical of winter, to spring when strong gradients exist in the SBC. The model results were checked against wind station time series, Special Sensor Microwave Imager wind speeds, and the European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis. The high-resolution (9-km grid) COAMPS wind shows expansion fans downwind of major capes where speed increases. The large-scale [O(100 km)] wind turns onshore in the Southern California Bight where both wind and wind stress curl weaken southward along the coast. The formation and evolution of the Catalina eddies are also simulated. These general features agree with observations. The turning appears to be the cumulative effect of synoptic cyclones shed downwind of Point Conception during periods of intense northerly wind. The turning and eddies are much weaker in the ECMWF reanalysis or the COAMPS field at the 81-km grid. Near the coast, observed small-scale (tens of kilometers) structures are reasonably reproduced by COAMPS at the 9-km grid. Results from the 9-km grid generally compare better with observations than the 27-km grid, suggesting that a more accurate model wind may be obtained at even higher resolution. However, in the SBC, simulated winds at both the 9- and 27-km grids show along-channel coherency during May, contrary to observations. The observed winds in the channel appear to be of small localized scales (≈<10 km) and would require an improved model grid and perhaps also boundary layer physics to simulate.

This study discusses the development of the Hurricane Forecast Improvement Program (HFIP) Corrected Consensus Approach (HCCA) for tropical cyclone track and intensity forecasts. The HCCA technique relies on the forecasts of separate input models for both track and intensity and assigns unequal weighting coefficients based on a set of training forecasts. The HCCA track and intensity forecasts for 2015 were competitive with some of the best-performing operational guidance at the National Hurricane Center (NHC); HCCA was the most skillful model for Atlantic track forecasts through 48 h. Average track input model coefficients for the 2015 forecasts in both the Atlantic and eastern North Pacific basins were largest for the European Centre for Medium-Range Weather Forecasts (ECMWF) deterministic model and the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS) ensemble mean, but the relative magnitudes of the intensity coefficients were more varied. Input model sensitivity experiments conducted using retrospective HCCA forecasts from 2011 to 2015 indicate that the ECMWF deterministic model had the largest positive impact on the skill of the HCCA track forecasts in both basins. The most important input models for HCCA intensity forecasts are the Hurricane Weather Research and Forecasting (HWRF) Model and the Coupled Ocean–Atmosphere Mesoscale Prediction System-Tropical Cyclone (COAMPS-TC) model initialized from the GFS. Several updates were incorporated into the HCCA formulation prior to the 2016 season. Verification results indicate HCCA continued to be a skillful model, especially for short-range (12–48 h) track forecasts in both basins.

PM2.5, a type of fine particulate with a diameter equal to or less than 2.5 micrometers, has been identified as a major source of air pollution, and is associated with many health issues. Research on utilizing various data sources, such as remote sensing and in situ sensors, for PM2.5 concentrations modeling remains a hot topic. In this study, the Next Generation Weather Radar (NEXRAD) is used as a supplementary weather data source, along with European Centre for Medium-Range Weather Forecasts (ECMWF), solar angles, and Geostationary Operational Environmental Satellite (GOES16) Aerosol Optical Depth (AOD) to model high spatial-temporal PM2.5 concentrations. PM2.5 concentrations as well as in situ weather condition variables are collected from the 31 sensors that are deployed in the Dallas Metropolitan area. Four machine learning models with different predictor variables are developed based on an ensemble approach. Since in situ weather observations are not widely available, ECMWF is used as an alternative data source for weather conditions in studies. Hence, the four established models are compared in three groups. Both models in this first group use weather variables collected from deployed sensors, but one uses NEXRAD and the other does not. In the second group, the two models use weather variables retrieved from ECMWF, one using NEXRAD and one without. In the third group, one model uses weather variables from ECMWF, and the other uses in situ weather variables, both without NEXRAD. The first two environmental groups investigate how NEXRAD can enhance model performances with weather variables collected from in situ observations and ECMWF, respectively. The third group explores how effective using ECMWF as an alternative source of weather conditions. Based on the results, the incorporation of NEXRAD achieves an R2 score of 0.86 and 0.83 for groups 1 and 2, respectively, for an improvement of 2.8% and 9.6% over those models without NEXRAD. For group three, the use of ECMWF as an alternative source of in situ weather observations results in a 0.13 R2 drop. For PM2.5 estimation, weather variables including precipitation, temperature, pressure, and surface pressure from ECMWF and deployed sensors, as well as NEXRAD velocity, are shown to be significant factors.

Despite recent advances in operational streamflow forecasting systems, anticipating and forecasting droughts and associated low-flow conditions remain major challenges in hydrology, with substantial impacts on water-dependent sectors such as agriculture and industry. Enhancing sub-seasonal to seasonal streamflow forecasting is therefore critical for improving water resources management. This study investigates the performance of a Handoff forecast Long Short-Term Memory (LSTM) architecture (Nearing et al., 2024) for probabilistic streamflow forecasting at lead times extending up to six months, with a particular emphasis on low-flow conditions.The Handoff forecast LSTM is trained regionally on a subset of 292 basins from the Catchment Attributes and MEteorology for Large-sample Studies - FRance dataset (CAMELS-FR) (Delaigue et al., 2025), after excluding basins affected by unreliable low-flow measurements. Model training relies on basin-averaged hydro-meteorological reanalysis data provided by CAMELS-FR. Evaluation of the model is conducted using ensemble streamflow forecasts generated from historical scenarios and meteorological ensemble predictions from the SEAS5 model from the European Center for Medium-Range Weather Forecasts (ECMWF) (Johnson et al., 2019)The generated ensemble streamflow forecasts are evaluated using a set of probabilistic metrics such as the Continuous Ranked Probability Score (CRPS), the Brier Score, the Area under the ROC curve, and the Talagrand diagram, and using the natural streamflow climatology as a reference. In addition, a sensitivity analysis of static catchment attributes is performed to assess their relative contribution to model performance and to better understand the drivers of predictability across basins.Delaigue, O., Guimarães, G. M., Brigode, P., Génot, B., Perrin, C., Soubeyroux, J.-M., Janet, B., Addor, N., & Andréassian, V. (2025). CAMELS-FR dataset: a large-sample hydroclimatic dataset for France to explore hydrological diversity and support model benchmarking. Earth System Science Data, 17(4), 1461–1479. https://doi.org/10.5194/essd-17-1461-2025Johnson, S. J., Stockdale, T. N., Ferranti, L., Balmaseda, M. A., Molteni, F., Magnusson, L., Tietsche, S., Decremer, D., Weisheimer, A., Balsamo, G., Keeley, S. P. E., Mogensen, K., Zuo, H., & Monge-Sanz, B. M. (2019). SEAS5: The new ECMWF seasonal forecast system. Geoscientific Model Development, 12(3), 1087–1117. https://doi.org/10.5194/gmd-12-1087-2019Nearing, G., Cohen, D., Dube, V., Gauch, M., Gilon, O., Harrigan, S., Hassidim, A., Klotz, D., Kratzert, F., Metzger, A., Nevo, S., Pappenberger, F., Prudhomme, C., Shalev, G., Shenzis, S., Tekalign, T. Y., Weitzner, D., & Matias, Y. (2024). Global prediction of extreme floods in ungauged watersheds. Nature, 627(8004), 559–563. https://doi.org/10.1038/s41586-024-07145-

Comparisons of PBL heights derived from CALIPSO and ECMWF reanalysis data over China

<p>Across the globe, there has been an increasing interest in improving the predictability of weekly to monthly (sub-seasonal) hydro-meteorological forecasts as they play a valuable role in medium- to long-term planning in many sectors such as agriculture, navigation, hydro-power production, and hazard warnings. A Precipitation-Runoff-Evapotranspiration HRU model (PREVAH) has been previously set up with raw metrological forcing of 51 ensemble members and 32 days lead time taken from the operational European Centre for Medium-Range Weather Forecasts (ECMWF) extended-range forecast. The PREVAH model is used to generate hydrological forecasts for the study area, which consists of 300 catchments covering approximately the entire area of Switzerland. The primary goal of this study is to improve the quality of the categorical forecast of weekly mean total discharge in a catchment laying in the lower, normal, or upper tercile of the climatological distribution at a monthly horizon. Therefore, we explore <span>the approach to post-process PREVAH outputs using machine learning algorithm Gaussian process</span>. Weather regime (WR) data, based on 500 hPa geopotential height in the Atlantic-European region are used as an added feature to further enhance the post-processing performance.</p><p>By comparing the overall accuracy and the ranked probability skill score of the post-processed forecasts with the ones of raw forecasts we show that the proposed post-processing techniques are able to improve the forecast skill. The degree of improvement varies by catchment, lead time and variable. The benefit of the added WR data is not consistent across the study area but most promising in high altitude catchments with steep slopes. Among the seven types of WRs, the majority of the corrections are observed when either a European blocking or a Scandinavian blocking is forecasted as the dominant weather regime. By applying a “best practice” to each individual catchment, that is the processing technique with the highest accuracy among the different proposed techniques, a median accuracy of 0.65 (improved from a value of 0.53 with no processing technique) can be achieved at 4-week lead time. Due to the small data size, the conclusions should be considered preliminary, but this study highlights the potential of improving the skill of sub-seasonal hydro-meteorological forecasts utilizing weather regime data and machine learning in a real-time deployable setup.</p>

Synoptic-scale vortices known as monsoon low pressure systems (LPSs) frequently produce intense precipitation and hydrological disasters in South Asia, so accurately forecasting LPS genesis is crucial for improving disaster preparedness and response. However, the accuracy of LPS genesis forecasts by numerical weather prediction models has remained unknown. Here, we evaluate the performance of two global ensemble models—the U.S. Global Ensemble Forecast System (GEFS) and the Ensemble Prediction System of the European Centre for Medium-Range Weather Forecasts (ECMWF)—in predicting LPS genesis during the years 2021–22. The GEFS successfully predicted about half the observed LPS genesis events 1–2 days in advance; the ECMWF model captured an additional 10% of observed genesis events. Both models had a false alarm ratio (FAR) of around 50% for 1–2-day lead times. In both ensembles, the control run typically exhibited a higher probability of detection (POD) of observed events and a lower FAR compared to the perturbed ensemble members. However, a consensus forecast, in which genesis is predicted when at least 20% of ensemble members forecast LPS formation, had POD values surpassing those of the control run for all lead times. Moreover, probabilistic predictions of genesis over the Bay of Bengal, where most LPSs form, were skillful, with the fraction of ensemble members predicting LPS formation over a 5-day lead time approximating the observed frequency of genesis, without any adjustment or bias correction.

The paper proposes a method for correcting the precipitation forecast of the European Centre for Medium-Range Weather Forecasts (ECMWF) model. The method first uses the ECMWF physical quantity field to draw the flow field diagram in the form of weather, and uses it as the input of the convolutional neural network(CNN) and performs feature extraction on it. Secondly, select several characteristic factors that are highly correlated with the forecast object from the many characteristic factors. Finally, the selected characteristic factors and the precipitation forecast factors of the ECMWF model are used as the input factors of the random forest regression model for modeling and forecasting. Through the correction forecast test of the next 24h precipitation at 10 forecast test stations, the results show that the ECMWF precipitation correction forecast method proposed in this paper is 15% less than the forecast method using ECMWF interpolation to the station on the MAE and RMSE indicators. And 17%. At the same time, on the TS scores of rainstorms above 24h, the revised forecast method proposed in this paper has a false alarm rate significantly lower than the forecast method of ECMWF interpolation to the station.