Meteorological Applications最新文献

{"title":"Interdisciplinary Assessment of Tanzanian Severe Weather Forecasts: Linking Verification and User Evaluation","authors":"Hellen E. Msemo,&nbsp;Cathryn E. Birch,&nbsp;Andrea L. Taylor,&nbsp;Tamora James,&nbsp;Andrew Hartley,&nbsp;Sarah Chapman,&nbsp;Andrew J. Dougill,&nbsp;Michael Baidu,&nbsp;Simon Peatman","doi":"10.1002/met.70220","DOIUrl":"https://doi.org/10.1002/met.70220","url":null,"abstract":"<p>While there is substantial research on weather forecast verification and on user evaluation of forecasts, these approaches are often treated separately. Linking them can enhance forecast credibility, increase trust, and support more effective decision-making. This study applies a mixed-methods approach to evaluate severe weather forecasts produced by the Tanzania Meteorological Authority (TMA), from forecast generation to end-user application. Quantitative verification used satellite-derived rainfall estimates from the Global Precipitation Measurement (GPM) mission to assess forecast skill across lead times and rainfall thresholds. Results show forecast accuracy above 0.80, with probability of detection (POD) of 0.6–0.9 across thresholds, but also a high false alarm ratio (0.4–0.7) and evidence of overforecasting, particularly at longer lead times. Reliance on a single national 50 mm/24 h threshold was a key driver of false alarms and may not reflect area-specific risks. Qualitative evaluation, based on interviews and participatory observations with disaster risk reduction stakeholders, highlighted that severe weather forecasts reached institutional actors such as the Dar es Salaam Emergency Response Team and the Tanzania Red Cross Society in a timely manner. However, challenges remain in ensuring last mile delivery to communities. Interviewees emphasized the importance of SMS, local radio, trusted intermediaries, and simpler language to improve understanding and uptake. The study recommends revising rainfall thresholds by regions, integrating disaster impact data into verification, and institutionalizing user feedback forums for severe weather forecasts. Limitations include reliance on 1 year of warnings, the absence of geo-referenced impact datasets, and constraints in polygon vectorization. By combining forecast verification with user evaluation, this study provides a comprehensive assessment of forecast performance and its application in disaster risk reduction, with lessons relevant beyond Tanzania.</p>","PeriodicalId":49825,"journal":{"name":"Meteorological Applications","volume":"33 5","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/met.70220","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149205718","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"The Influence of Sudden Stratospheric Warmings on Extreme European Electricity Demand","authors":"Regan Mudhar,&nbsp;William J. M. Seviour,&nbsp;Hannah C. Bloomfield,&nbsp;Amy H. Butler,&nbsp;Chaim I. Garfinkel,&nbsp;Peter Hitchcock,&nbsp;James A. Screen,&nbsp;Stephen I. Thomson","doi":"10.1002/met.70241","DOIUrl":"https://doi.org/10.1002/met.70241","url":null,"abstract":"<p>Sudden stratospheric warmings (SSWs) provide a source of surface predictability up to several weeks in advance. They generally precede cold weather across the northern midlatitudes, including extremes. When it is cold, demand for heating typically increases electricity demand. We use reanalysis data and a weather-driven demand model for various European countries to investigate the relationship between SSWs and extreme electricity demand. We find that extreme demand of ~120% of climatology is 1.5–3 times more likely in periods following SSWs across northern and central Europe. We then use subseasonal forecast experiments from the Stratospheric Nudging and Predictable Surface Impacts project to explicitly attribute the impact of the February 2018 SSW, an event that preceded a cold weather outbreak across Europe. These experiments include nudging of the stratospheric state to either its observed or climatological evolution, allowing the causal impact of the SSW to be identified. We find a robust simulated increase in extreme demand risk attributed to the 2018 event. This indicates a window of opportunity for earlier prediction of extreme European electricity demand risks following SSWs.</p>","PeriodicalId":49825,"journal":{"name":"Meteorological Applications","volume":"33 5","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/met.70241","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149146228","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Joint Forecasting of Wind Speed and Direction via Ensemble Post-Processing","authors":"Alice Lake,&nbsp;Matthew Fry,&nbsp;Stephen Moseley,&nbsp;Alasdair Skea","doi":"10.1002/met.70254","DOIUrl":"https://doi.org/10.1002/met.70254","url":null,"abstract":"<p>Wind speed and direction are often considered jointly in operational decision-making, particularly in sectors such as renewable energy and aviation, where physically-consistent forecasts are required to support operational planning and safety-critical decisions. Many such users require deterministic-style forecasts, despite ensemble forecasting increasingly becoming the standard approach taken by weather forecasting centres. Current operational practice is to provide single-value forecasts by delivering the control (unperturbed) member of the ensemble, but this approach fundamentally undermines the purpose of the ensemble since it ignores information from all other members. This paper presents an alternative approach by extending a recently-proposed ensemble post-processing technique to produce physically-consistent spot forecasts of wind speed and direction. Predictive probability density functions are estimated using a Gamma kernel for wind speed and a von Mises kernel for wind direction, representing the linear and circular properties of the variables respectively. A single-value forecast is then obtained by selecting the ensemble member which maximises its likelihood under the joint densities, ensuring physical consistency across variables while making use of information from the entire ensemble. The method is applied to the operational Met Office convective-scale ensemble and evaluated over a full calendar year (2024), using both surface synoptic observations and radiosonde measurements at 100 m and 1000 m AGL. Results show that maximising likelihood across the full 5-day forecast horizon produces forecasts broadly comparable to the operational control member, used here as a baseline to reflect current operational practice. A piecewise 24-h selection strategy, however, demonstrates the potential to improve upon the control member, particularly at longer lead times. A case study of Storm Darragh at two major UK airports further demonstrates the practical value of the approach during a high-impact event, where the likelihood-based selection provided earlier and more coherent guidance on hazardous wind conditions than the operational control member.</p>","PeriodicalId":49825,"journal":{"name":"Meteorological Applications","volume":"33 5","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/met.70254","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149145931","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Evaluation of Satellite Rainfall Products for Nowcasting Over Southern Africa Using Rain-Gauges and Lightning Data","authors":"Simon Ageet,&nbsp;John H. Marsham,&nbsp;Douglas J. Parker,&nbsp;Abhilash Singh,&nbsp;Christopher M. Taylor,&nbsp;Alexander J. Roberts,&nbsp;James Groves,&nbsp;Edson Nkonde,&nbsp;Adrià Amell","doi":"10.1002/met.70253","DOIUrl":"https://doi.org/10.1002/met.70253","url":null,"abstract":"<p>Satellite-based ‘nowcasting’, which extrapolates the latest observations to provide forecasts from minutes to 6 h ahead, offers an accurate and timely approach for monitoring extreme weather. We evaluate the widely used NWCSAF Convective Rainfall Rate (CRR) nowcasting product using instantaneous (non-propagated) estimates against a network of high-temporal-resolution rain gauges in Zambia during the wet seasons from 2019 to 2024. We also assess two alternative satellite-based rainfall products: IMERG microwave-only (IMERG-HQ) and Rain over Africa (RoA), a machine learning–generated nowcasting product. Additionally, lightning data from the Meteosat Third Generation mission and cloud-top temperature data were used to evaluate the products. These datasets formed part of the products used by forecasters during nowcasting testbeds conducted under the WISER EWSA project and positive feedback from users was widely reported. Our analysis shows that the products have distributions similar to the in situ data, but all exhibit biases, with CRR and IMERG-HQ generally overestimating rainfall rates while RoA tends to underestimate. Across different thresholds and aggregation windows, RoA outperforms IMERG-HQ and CRR for rain occurrence on most metrics, including hit rate, probability of false alarm and Heidke skill score. However, all products show reduced skill for extreme rainfall. CRR also shows weak correspondence with lightning activity and does not detect rainfall from cloud tops warmer than −30°C. In contrast, RoA and IMERG-HQ exhibit greater sensitivity to cloud top temperature, including warmer cloud regimes. The generally superior performance of RoA compared to IMERG-HQ is remarkable, given that RoA only takes infrared satellite channels as input, whereas IMERG-HQ uses microwave channels. We attribute RoA's greater skill to the ability of ML to optimally exploit the data and bring spatial information into the algorithm. Our findings emphasise the value of satellite-based monitoring and nowcasting, while underscoring the need for training and calibration to improve accuracy.</p>","PeriodicalId":49825,"journal":{"name":"Meteorological Applications","volume":"33 5","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/met.70253","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149145772","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Extensions to Postprocessing of Hourly Ensemble Probability of Precipitation Using a Neural Network With Temporally Linked Predictors","authors":"Martin Davis Apsitis","doi":"10.1002/met.70255","DOIUrl":"https://doi.org/10.1002/met.70255","url":null,"abstract":"<p>Forecasting the hourly probability of precipitation (PoP) remains a persistent challenge, particularly across varying precipitation regimes and at fine temporal resolution. This study evaluates whether a recently proposed machine-learning-based postprocessing framework based on the use of available raw control forecast predictors in complementing the ensemble predictions can be robustly extended to a larger spatial domain and a full annual cycle. Using the European Centre for Medium-Range Weather Forecasts (ECMWF) ensemble and control forecast data at 0.1° resolution, probabilistic precipitation forecasts are generated for 25 stations in Latvia over the period from 1 March 2023 to 29 February 2024. Three supervised machine-learning classifiers—logistic regression, random forest, and a neural network—are trained daily to produce baseline (‘Base’) PoP predictions. Predictor selection is performed using a custom cross-validation scoring function designed to avoid probability collapse in rare-event regimes. Across most stations, lead times, and seasons, all Base methods exhibit lower skill than the raw ECMWF ensemble, as assessed by Brier skill score (BSS), Brier score (BS) timelines, receiver characteristic operation (ROC) curves, and contingency metrics. Apparent improvements during summer months are primarily associated with systematic under-forecasting in low-frequency precipitation regimes rather than enhanced discrimination. Motivated by temporal persistence observed in hourly precipitation, a secondary postprocessing step is introduced. A neural network trained on sliding temporal windows of Base predictions and raw ensemble predictors produces revised (‘Post’) forecasts. Post predictions demonstrate improved calibration, reduced false alarm ratios, and positive BSS values for most stations, particularly during drier months. However, the overall gains are modest and largely reflect probability dampening rather than consistent improvements in event detection. These results suggest that raw control-forecast predictors, when used without prior postprocessing, do not necessarily provide useful information beyond the raw ensemble.</p>","PeriodicalId":49825,"journal":{"name":"Meteorological Applications","volume":"33 5","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/met.70255","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149145762","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Turning up the Heat: Assessing 2-m Temperature Forecast Errors in Artificial Intelligence Weather Prediction Models During Heat Waves","authors":"Kelsey E. Ennis,&nbsp;Marybeth C. Arcodia,&nbsp;Martin A. Fernandez,&nbsp;Eric D. Maloney,&nbsp;Elizabeth A. Barnes","doi":"10.1002/met.70249","DOIUrl":"https://doi.org/10.1002/met.70249","url":null,"abstract":"<p>Extreme heat is the deadliest weather-related hazard in the United States. Furthermore, heat extremes are increasing in intensity, frequency, and duration, making skillful forecasts vital to protecting life and property. Traditional numerical weather prediction (NWP) models struggle with extreme heat forecasts on longer (i.e., medium-range- and subseasonal) timescales. Meanwhile, artificial intelligence-based weather prediction (AIWP) models are progressing rapidly. However, it is largely unknown how well AIWP models predict extreme heat, especially for longer-range forecasts. This study investigates 2-m temperature forecasts for 60 heat waves across four contiguous United States regions (Northwest, Midwest, Northeast, Southeast) and the four boreal seasons at medium-range to subseasonal (i.e., 10–20-day) lead times, using two deterministic AIWP models (GraphCast and Pangu-Weather) and one probabilistic AIWP model (GenCast). All three models exhibit persistent mean cold biases throughout the forecast evolution, especially near the start of the heat wave. However, the GenCast ensemble mean consistently exhibits smaller biases and errors than the two deterministic models. In all three AIWP models, the smallest regional average errors during the heat wave period are in the Northwest, with the largest in the Midwest. In all boreal seasons, mean cold biases occur before the heat waves in the three AIWP models except Pangu-Weather in winter, which exhibits a mean warm bias before heat wave onset. A brief comparison of 2-m temperature forecast errors is then illustrated among the AIWP models and a traditional NWP model (NOAA United Forecast System Global Ensemble Forecast System (UFS GEFS)), for which GenCast particularly performs strongly. Overall, results show that the probabilistic AIWP model GenCast offers the most promising accuracy toward medium-range and subseasonal forecasts of extreme heat, albeit with a consistent slight mean cold bias.</p>","PeriodicalId":49825,"journal":{"name":"Meteorological Applications","volume":"33 5","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-09-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/met.70249","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149146994","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"How Far Can Meteorological Stations Provide Reliable Data for Reference Evapotranspiration Estimation?","authors":"Rahman Barideh,&nbsp;Fereshteh Nasimi,&nbsp;Alireza Tavasolee","doi":"10.1002/met.70251","DOIUrl":"https://doi.org/10.1002/met.70251","url":null,"abstract":"<p>Accurate estimation of reference evapotranspiration (RET) is essential for water resource management in water-stressed regions such as the Urmia Lake basin, northwestern Iran. While RET can be readily estimated from station records, the spatial extent over which a single station's data remain reliable is seldom quantified. This study addresses this gap by developing a spatial accuracy-radius framework that defines, for each of 24 synoptic meteorological stations in the Urmia Lake basin, the geographic extent within which its FAO Penman-Monteith RET estimates remain reliable, using ERA5-Land gridded data (463 points, 9 km resolution, 2010–2020) as a spatial reference. Station accuracy was assessed using the normalized root mean square error (nRMSE), and station-specific accuracy radii were delineated based on the station with the minimum nRMSE at each point. The results show that station accuracy decreases with increasing distance from the station, and that stations located in urban and industrial areas exhibit lower reliability, consistent with the urban heat island effect and other local environmental influences. Based on this analysis, 18 of the 24 stations were identified as valid and used to define the accuracy-radius coverage of the basin. Compared with the conventional nearest-station approach, the proposed accuracy-radius method increased the mean basin area with nRMSE below 0.2 from 70% to 91%, and reduced the mean basin-wide nRMSE from 0.197 to 0.168. These findings indicate that accounting for station-specific spatial reliability, rather than proximity alone, provides a more robust basis for meteorological station selection in hydrological modeling, agricultural planning, and sustainable water resource management.</p>","PeriodicalId":49825,"journal":{"name":"Meteorological Applications","volume":"33 5","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-09-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/met.70251","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149146541","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"An Arctic Field Campaign Using Stratospheric Balloons for Micro-Dropsonde Deployment: Insights and Evaluation","authors":"Matthew Fry,&nbsp;Dawn Harrison,&nbsp;Jonny King","doi":"10.1002/met.70244","DOIUrl":"https://doi.org/10.1002/met.70244","url":null,"abstract":"<p>The Arctic remains one of the most data-sparse regions on Earth, limiting the accuracy of numerical weather prediction and operational forecasting. A February 2025 field campaign, conducted in Iceland, tested a prototype stratospheric dropsonde system aimed at enhancing vertical atmospheric profiling across the Arctic environment. The campaign intended to demonstrate the systems' capability to operate in harsh Arctic environments, as well as its potential to address observational gaps in the troposphere critical for improving forecasts of high-impact weather systems such as Arctic cyclones. Over a 12-day period, 10 Voltitude StratoSondes were launched from Eastern Iceland, subsequently releasing targeted micro-dropsondes from a range of Arctic locations. These data are compared against the Met Office Global Model to assess biases in temperature, humidity and wind speed/direction. Averaged across 40 dropsonde descents, mean biases with respect to the Met Office's global model were found to be +0.32°C for temperature, −3.9% for relative humidity and −0.22 and +0.06 m s⁻¹ for zonal and meridional wind components, respectively. Preliminary results indicate that micro-dropsonde profiles provide valuable high-resolution observations in regions poorly sampled by conventional networks. The campaign also highlights logistical and technical challenges associated with Arctic field operations and underscores the potential for emerging balloon-based observation systems to address critical polar observation gaps.</p>","PeriodicalId":49825,"journal":{"name":"Meteorological Applications","volume":"33 5","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-09-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/met.70244","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149048038","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Analysis of Low-Level Windshear Cases at the Hong Kong International Airport Under Prevailing Easterly Wind Situations","authors":"Man Lok Chong,&nbsp;Ping Cheung,&nbsp;Kai Kwong Lai,&nbsp;Pak Wai Chan","doi":"10.1002/met.70246","DOIUrl":"https://doi.org/10.1002/met.70246","url":null,"abstract":"<p>During spring, low-level windshear occurs at the Hong Kong International Airport (HKIA) under various easterly wind conditions associated with low-level jets, terrain and sea breezes. This paper analyses in detail several cases of low-level windshear by considering pilot reports, aircraft data and measurements from Doppler Light Detection and Ranging (LIDAR) systems. The LIDARs are found to provide crucial data for capturing windshear features related to terrain, low-level jet and synoptic forcing. Their glide-path scan data are highly useful for constructing the headwind profiles, which are used to generate windshear alerts. The LIDAR-based headwind profiles are consistent with pilot reports and aircraft data for a variety of windshear cases. Through case studies, this paper also demonstrates the subjective nature of pilot windshear reports, and the transient and sporadic nature of terrain-induced windshear. In a significant windshear case with 16 pilot reports, the capability of a high-resolution numerical weather prediction (NWP) model to provide windshear alerts is evaluated. The model is found to be skilful in capturing the pilot reports through relative operating characteristics (ROC) curves, though the results should be interpreted with the limited sample size in mind. This paper shows the usefulness of LIDAR in providing detection-based windshear alerts and the potential capability of a high-resolution NWP model in providing forecast windshear alerts with a lead time of more than 24 h.</p>","PeriodicalId":49825,"journal":{"name":"Meteorological Applications","volume":"33 5","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-09-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/met.70246","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149048037","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Extended-Range Forecasting Experiments Basedon Multi-Intelligent Models for Regional Precipitation Process","authors":"Lingling Zeng,&nbsp;Xianyun Wu,&nbsp;Bin Zhu,&nbsp;Guirong Tan,&nbsp;Hui Zhao,&nbsp;Yi Zhang","doi":"10.1002/met.70245","DOIUrl":"https://doi.org/10.1002/met.70245","url":null,"abstract":"<p>Based on the circulation products from European Centre for Medium-Range Weather Forecasts (ECMWF) model and daily site rainfall data over Hunan Province during May to August between 2005 and 2022, experiments for extended precipitation forecasting were conducted by incorporating the Convolutional Block Attention Module (CBAM) and Long Short-Term Memory Network (LSTM) into the Convolutional Neural Network (CNN) separately. The performance of the CNN, CNN-CBAM, CNN-LSTM, and ECMWF ensemble was evaluated and compared. The results indicate that the forecasting skill of ECMWF ensemble significantly decreases as the forecast horizon extends. In contrast, the skill of the three intelligent models (CNN, CNN-CBAM, and CNN-LSTM) exhibits greater stability. Notably, the CNN-LSTM model maintains high forecasting skill even at the 20–30 day forecast horizons. Furthermore, the superiority of the intelligent models becomes increasingly evident with higher precipitation magnitudes. For instance, while the ECMWF ensemble shows nearly no skill for larger precipitation events (daily rainfall exceeding 25 mm), the intelligent models retain a certain level of consistent accuracy. Case studies highlight that ECMWF only holds a slight advantage in predicting light precipitation, whereas the intelligent models outperform ECMWF for heavier precipitation events. In particular, the CNN-LSTM model demonstrates superior performance in long-range forecasting.</p>","PeriodicalId":49825,"journal":{"name":"Meteorological Applications","volume":"33 5","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-09-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/met.70245","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148994197","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}