Abstract:Using the operational practices of the National Meteorological Center as a case study, this paper critically examines the technical limitations currently faced in short-term and medium- to long-term weather forecasting as efforts progress seamless forecasting capabilities. Key challenges include improving the accuracy of forecasts for extreme hazardous weather events and transitioning toward intelligent forecasting systems that effectively integrate both subjective and objective methods. Building on the historical trajectory of meteorological forecasting—particularly the advancement of numerical weather prediction through the integration of atmospheric, computational, and information sciences—this study highlights the growing importance of artificial intelligence (AI) as a critical enabler of forecasting capability. The article provides a comprehensive review of AI applications in meteorology over the past decade, including its use in monitoring and forecasting severe convective weather, typhoons, quantitative precipitation, secondary disasters such as mountain floods, intelligent grid-based forecasting of meteorological variables, risk assessments of meteorological disasters, and the automated generation of text and graphic forecast products. Particular attention is given to two AI-based models released in 2024: Fenglei, a nowcasting system for high-impact weather, and Fengqing, a global short- and medium-term forecasting model. The integration of AI with traditional meteorological techniques has significantly improved the accuracy of forecasts for typhoons, heavy rainfall, and severe convection. The deployment of AI across operational systems at the National Meteorological Center has accelerated the digital and intelligent transformation of weather forecasting. The release and application of the Fenglei and Fengqing models represent a substantial step forward, positioning the center at the forefront of meteorological forecasting innovation. Nonetheless, several urgent issues remain from an operational perspective. There is a crucial need to reduce dependency on high-quality foreign datasets by leveraging the China Meteorological Administration's observational capabilities to build long-term, high-quality meteorological datasets. To address the inherent “black box” nature of AI models, effective statistical tools must be developed for interpreting their predictions. Moreover, integrating forecasters' intuitive knowledge into AI systems may enhance model performance in forecasting rare extreme weather events, for which training samples are limited. Finally, advancing intelligent forecasting operations requires the development of integrated AI systems capable of supporting meteorologists across all phases of forecasting. This includes the construction of digital AI forecasting assistants that combine various algorithms—such as numerical models, intelligent grid forecasts, and large language models—into a unified operational framework.

Abstract:Subseasonal prediction—forecasting weather and climate phenomena 2 to 6 weeks ahead—plays a pivotal role in sectors ranging from agriculture and disaster risk reduction to water resource management and energy planning.Accurate predictions at this timescale are critical for mitigating impacts of extreme events like floods,droughts,and heatwaves,yet they remain notoriously challenging.Traditional numerical weather prediction (NWP) models,despite advancements through ensemble systems,suffer from diminishing predictability due to rapid decay of initial condition signals.Machine learning (ML) approaches offer promise but have shown limited success in subseasonal scales,hindered by narrow variable coverage,insufficient uncertainty quantification,and reliance on foreign datasets like ERA5,which poses risks to data sovereignty and operational autonomy.To address these gaps,this study introduces “Fengshun” (CMA-AIM-S2S-Fengshun),an artificial intelligence (AI) large model developed collaboratively by the National Climate Center and Fudan University.Leveraging domestically produced CRA-40 reanalysis data and FY-3E satellite observations,the model aims to establish a robust,autonomous subseasonal prediction framework for China's regional precipitation.The “Fengshun” system employs a cascaded Swin Transformer architecture to model spatiotemporal dependencies in meteorological fields.Its core innovation lies in a novel intelligent perturbation generation module,which integrates Kullback-Leibler (KL) divergence and L1 loss optimization to learn low-rank Gaussian distributions of historical data and prediction time features.During inference,this module generates probabilistic ensemble forecasts by sampling perturbation vectors,effectively mitigating error accumulation in autoregressive predictions and providing probabilistic representations of future climate states.Unlike many existing AI models dependent on foreign datasets,“Fengshun” is trained entirely on Chinese-controlled data,ensuring real-time data assimilation (with same-day updates,5 days faster than ERA5-based models) and operational independence.Historical hindcasts from 2017 to 2021 (independent of the training dataset) were validated against ECMWF's subseasonal-to-seasonal (S2S) predictions,using CRA-40 reanalysis as ground truth.The evaluation focused on precipitation anomaly percentage over China,with metrics including temporal correlation coefficient (TCC) for skill assessment and root-mean-square error (RMSE) for magnitude accuracy.while case studies examined the model's performance during the July 2024 North China heavy precipitation event.“Fengshun” outperformed ECMWF across most subseasonal ranges (15—45 days lead time),achieving an 18.6％ improvement in TCC and a 7.8％ reduction in RMSE for national-averaged pentad precipitation.The model demonstrated exceptional skill in South China (41.2％ TCC improvement) and East China (26.5％ improvement),maintaining predictability up to 8 pentads (40 days) during the summer flood season,a period critical for disaster preparedness.The Madden-Julian Oscillation (MJO),a key driver of subseasonal variability,was predicted with a skill retention time of 32 days using CRA-40 data,surpassing ECMWF's 30-day benchmark.This advancement is attributed to the model's ability to capture tropical-extratropical interactions,which are fundamental to East Asian monsoon dynamics and precipitation patterns.For the mid-July 2024 event,“Fengshun” accurately predicted the spatial distribution and intensity of precipitation anomalies 3—4 pentads (15—20 days) in advance.At a 7-day lead time,its area correlation coefficient (ACC) reached 0.80,compared to ECMWF's 0.68,and threat scores (TS) for extreme precipitation (>100％ anomaly) were 0.36 versus ECMWF's 0.24,highlighting superior early-warning capabilities.By leveraging indigenous data and advanced AI architecture,“Fengshun” delivers robust subseasonal precipitation forecasts with marked improvements over traditional models,particularly in regions vulnerable to monsoon-driven extremes.Its operational deployment promises to enhance proactive disaster response,agricultural planning,and water resource management,exemplifying the potential of AI to transform climate science and service.

Abstract:With the intensification of global warming,extreme heatwave events are occurring with increasing frequency and intensity,posing severe impacts to human society and ecosystems.Accurate prediction of extreme heatwaves is essential for urban resilience,as megacities are particularly vulnerable due to impacts on public health,energy supply,and transportation.Sub-seasonal prediction,which bridges short-term weather forecasting and seasonal prediction,plays a critical role in mitigating the effects of extreme heatwaves.However,traditional sub-seasonal prediction studies have primarily focused on temperature anomalies or their low-frequency components.Recent advancements in meteorological artificial intelligence (AI) models have led to the rapid developments in weather prediction,but their capability for sub-seasonal heatwave prediction in megacities remains unclear.This study evaluates the performance of three AI-driven meteorological models (Pangu,FuXi,and FourCastNet) in predicting heatwave events in Shanghai during midsummer 2024.The evaluation is based on correlation skill,power spectrum analysis,and other diagnostic methods.Analysis of heatwaves and their associated circulation patterns indicates that the number of high-temperature days in Shanghai during midsummer is significantly and positively correlated with the strength of the subtropical high aloft.The timing of heatwave occurrences aligns closely with the positive phases of the 10—20 d quasi-periodic oscillation and is simultaneously influenced by the 30—60 d low-frequency oscillation.On the quasi-biweekly timescale,extreme heatwave events in Shanghai are linked to both the northwestward propagation of wave trains triggered by convective anomalies in the tropical western Pacific and the eastward propagation of circulation anomalies along the Silk Road teleconnection pattern in the middle and high latitudes.Among the three AI models,both Pangu and FuXi demonstrate skillful high-temperature predictions up to 15 d in advance.Certain models,such as Pangu,can effectively predict the evolution of the subtropical high associated with heatwaves at lead times of 16—20 d.Differences in sub-seasonal prediction performance among the AI models are associated with their ability to capture low-frequency evolution of high temperatures and atmospheric circulation.Models with stronger predictive capabilities for low-frequency circulation anomalies tend to produce more accurate sub-seasonal forecasts of the subtropical high and high-temperature events.Furthermore,AI models exhibit greater predictive skill for sub-seasonal heatwaves during periods not affected by the end of Meiyu season or typhoon activity.This study also highlights that the differences in sub-seasonal prediction skill among the three AI models for low-frequency circulation at lead times of 11—15 d are influenced by the intensity of the initial low-frequency state.The uncertainty in prediction skill caused by initial-state variability warrants further investigation.Given that this study evaluates AI model performance primarily using 2024 as a case study,more comprehensive assessments are needed.Additionally,when applying AI models to enhance sub-seasonal heatwave prediction in megacities,the influence of surface heterogeneity,such as urbanization effects,should be considered to improve forecast accuracy.

Abstract:Clouds play a vital role in the earth-atmosphere system.Accurate cloud classification is essential for improving regional weather forecasts and understanding the global energy budget.However,precise and objective identification of ground-based cloud images remain challenging,primarily due to the limited availability of standardized cloud image datasets.This constraint hampers the further development of deep learning models for cloud classification.To address this issue,we propose a methodological hypothesis:can pre-training deep learning models on large scale,non-meteorological datasets enhance the accuracy of cloud classification,followed by fine-tuning with domain-specific cloud imagery? To test this hypothesis,we implement three deep learning architectures—two convolutional neural networks (ResNet50 and MobileNet-V2) and a self-attention-based Vision Transformer (ViT)—to perform ground-based cloud classification.We conduct a comparative analysis of models trained solely on cloud image datasets and those pre-trained on the ImageNet dataset before being fine-tuning with cloud data.Our results highlight the impact of pre-training strategies across different architectures.Even without pre-training,ResNet50 and MobileNet-V2 achieve strong baseline performance,with average F₁ scores of 0.85 and 0.87,respectively.Notably,the ViT model shows significant improvement with pre-training:the F₁ score increases from 0.79 to 0.96—a 21.5％ enhancement—demonstrating the importance of large-scale pre-training for architectures reliant on spatial feature extraction.Analysis of misclassified cases reveals that deep learning models primarily rely on spatial characteristics to distinguish cloud types.This suggests that incorporating auxiliary meteorological parameters—such as cloud-base height and thickness—as embedded features may further enhance model interpretability and performance.The performance gains from pre-training are largely attributed to improved edge detection and morphological pattern recognition,which are especially beneficial for complex architectures like ViT.In addition to these theoretical contributions,this study achieves practical implementation by deploying a stable cloud classification model on a mobile platform (available at http://43.142.162.19:5174/).This application supports real-time cloud-type identification via photo uploads and provides educational content,thereby promoting public engagement with atmospheric science and demonstrating the real-world applicability of deep learning for ground-based cloud observation.
Future research will focus on integrating cloud physical properties—such as thermodynamic parameters and radiative characteristics—into deep learning models.Fusing physical constraints with visual features may enhance classification robustness,reduce data requirements,and improve interpretability,paving the way for explainable AI systems in atmospheric sciences.In conclusion,this study establishes deep learning as an effective approach for automated cloud classification and underscores the critical role of pre-training,especially for advanced architectures.The mobile deployment further bridges meteorological research and public outreach,demonstrating the dual scientific and educational value of AI-powered cloud classification systems.

Abstract:The planetary boundary layer (PBL),located in the lower troposphere near the earth's surface,is profoundly influenced by surface friction,thermal processes,and evaporation.As a crucial component of the atmospheric system,the PBL acts as a bridge between the free atmosphere and the Earth's surface and serves as the primary space for human activity.Planetary boundary layer height (PBLH),a key structural characteristic,reflects physical processes such as turbulent mixing and convective development within the boundary layer.Accurately tracking its continuous changes and evolution is essential for advancing research in atmospheric science,environmental monitoring and pollution control.
Traditional methods for PBLH determination,such as sounding observations,offer high accuracy but are limited in spatial and temporal coverage,restricting their utility for multi-scale continuous observations.Remote sensing can provide continuous monitoring but is significantly affected by weather conditions and cannot fully capture PBLH dynamics.Numerical models,while useful,are subject to intrinsic model errors.A need remains to further investigate the relationship between near-surface atmospheric characteristics and PBLH.In this study,we apply a machine learning approach,XGBoost,to predict PBLH using long-term surface meteorological,wind radar,and sounding data from Beijing (January 2016 to May 2019) to train a model,which we subsequently employ to predict PBLH from June 2019 to May 2020.
Results indicate that the model performs optimally under clear-sky daytime conditions,achieving a high correlation with radiosonde-derived PBLH (correlation coefficient=0.86).Prediction accuracy is reduced at night.Surface temperature,relative humidity,and wind speed emerge as the most influential input features.The predicted PBLH displays a pronounced diurnal cycle,increasing rapidly after sunrise,gradually decreasing in the afternoon,and stabilizing at night.Seasonal analysis shows that daily PBLH variations are more pronounced in spring and summer,reaching up to 1 km,and are smaller in autumn and winter,around 700 m.
Overall,the XGBoost algorithm outperforms multiple linear regression and support vector regression in PBLH predictions,offering an efficient,intuitive method to continuously estimate PBLH's diurnal variation.This approach provides new insights into the diurnal and seasonal patterns of the PBL,supporting multi-period analysis.However,model performance for nighttime PBLH is limited,as it does not fully capture the stabilized boundary layer's vertical stricture due to strong radiative cooling and the weakened interaction between the PBL and the surface.Future work will incorporate vertical observation data to refine the model structure and compare results with other detection methods to validate the applicability of the XGBoost algorithm.

Abstract:Gusts significantly impact the safety of maritime shipping and offshore operations.However,due to the complex mechanisms driving wind speed variability,accurate gust prediction remains a longstanding challenge.To improve the forecasting accuracy of offshore gusts,this study utilized hourly maximum wind speed observations from the China Meteorological Administration (January 2021 to December 2022) along with 24-hour forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF) deterministic model.Based on data from 15 buoy stations located in China's coastal waters,three distinct gust forecasting methods were developed:1) a model based on a fully connected neural network (FCNN),2) a model applying quantile matching to the 10-m wind speed forecasts from the numerical model,and 3) a hybrid model combining FCNN prediction with subsequent quantile matching correction (FCNN+QM).
These methods were compared and validated using independent data from January to December 2023,leading to the following conclusions:When used alone,the FCNN method tends to significantly underestimate strong gust events.To address this,the study explored transforming the prediction target from gust wind speed to the adjustment value between gusts and the 10-m wind speed (or wind speeds at other pressure levels) from the numerical model.Although this transformation improved the distribution of the dependent variable,making it closer to normal and alleviating sample imbalance issues,it also introduced additional errors from the numerical model into the dependent variable.Comparative experiments demonstrated that modifying the prediction target alone could not fully compensate for the FCNN's limitations in capturing extreme wind speeds.Applying a secondary correction through quantile matching (FCNN+QM) substantially improved the model's performance for strong gusts while maintaining stable accuracy for weaker winds.Moreover,direct quantile matching of the 10-m wind speed forecasts from ECMWF produced prediction skill comparable to,or even exceeding,that of the FCNN+QM method for strong gusts.This finding suggests that,in the context of offshore gust prediction,the FCNN step is not strictly necessary when using quantile matching.Nevertheless,incorporating FCNN as a preliminary step enhances the robustness of strong gust prediction overall.Validation results across 2023 confirmed that the FCNN+QM method achieved the best performance for strong gust forecasts.
The combined FCNN and quantile matching approach,trained uniformly across 15 coastal buoy stations,demonstrated strong applicability for offshore gust prediction across diverse sea areas.However,slight overestimation of strong gusts was observed at a few individual stations.Given the scarcity of offshore observations and the complexity of influencing meteorological systems,this generalizable model provides a valuable reference for poorly observed regions,as demonstrated using during Typhoon Doksuri's strong wind event.While the FCNN approach enables more detailed quantification of the influence of different meteorological variables on gusts,its predictive capability remains limited by the accuracy of the underlying numerical model and the variability of offshore weather systems.To further enhance offshore gust forecasting,future work should explore segmented machine learning modeling tailored to specific conditions,combined with quantile matching correction,to achieve more targeted and accurate predictions.

Abstract:The El Ni?o-Southern Oscillation (ENSO) is the most significant interannual climate variability phenomenon, exerting profound influences on global weather patterns and climate anomalies. The associated natural disasters pose severe threats to human lives and property. Traditional ENSO prediction methods primarily include dynamical and statistical approaches. Due to the long-term accumulation of climate data and a well-established theoretical foundation of ENSO dynamics, these methods have been extensively developed. Studies have shown that traditional methods perform well within the first 6 months of forecasting, achieving a correlation coefficient skill (Corr) of up to 0.85. However, prediction accuracy declines over time, with most models struggling to maintain a Corr above 0.5 beyond 12 months. This limitation is largely attributed to the inherent nonlinearity and uncertainty of ENSO events, which challenge the ability of traditional models to improve prediction accuracy and extend forecast lead times. Additionally, computational error accumulation, empirical limitations, and uncertainties in parameter optimization restrict the effectiveness of dynamical models for key long-term ENSO prediction. Likewise, due to the highly nonlinear nature of ENSO onset and evolution, statistical models struggle to capture the complex intrinsic features of ENSO from large datasets, thereby limiting prediction accuracy.
In recent years, deep learning techniques have garnered increasing attention in ENSO forecasting due to their ability to efficiently process complex spatiotemporal data and adaptively learn feature representations. Researchers have explored deep learning approaches for ENSO prediction, achieving promising results. This review provides a comprehensive discussion of ENSO prediction, beginning with an overview of ENSO-related knowledge, including key datasets for ENSO classification and forecasting. It then examines traditional ENSO prediction methods, covering both dynamical and statistical approaches. The review further explores the application of deep learning models in ENSO forecasting, including methods based on convolutional neural networks (CNNs), recurrent neural networks (RNNs), graph neural networks (GNNs), and Transformer models. The advantages, limitations, and development trends of each type approach are summarized.
Despite the promising advancements in deep learning for ENSO prediction, several key challenges remain: 1) The “black-box” nature of deep learning models limits the physical interpretability of predictions. Although efforts have been made to integrate physical knowledge with deep learning, research on model interpretability remains incomplete. 2) The limited time span of ENSO observational data and the rarity of extreme ENSO events result in constrained training samples. Additionally, discrepancies between simulated and observed data pose challenges, necessitating further exploration of multivariate information to enhance model performance. 3) The ongoing rapid changes in global climate may alter ENSO characteristics, making deep learning models trained on historical data susceptible to reduced reliability. Incorporating climate change impacts into deep learning models is essential for improving forecast robustness.

Abstract:Subseasonal temperature anomaly prediction is a critical component of short-term climate forecasting in China. Accurate forecasts are essential for effective response to meteorological hazards such as heatwaves and cold surges. In recent years, artificial intelligence (AI) has been increasingly applied in climate prediction, enabling the detection of predictable signals related to extreme climate events and the development of statistical prediction models. However, many existing approaches underutilize valuable information from dynamical models. To address this limitation, this study proposes a deep learning framework that integrates dynamical model outputs to improve subseasonal temperature anomaly prediction. Using daily atmospheric circulation data from the NCEP/NCAR reanalysis, outgoing longwave radiation (OLR) data from NOAA, and daily temperature records from 97 national weather stations in Hunan Province for the period 1981—2023, we develop a pre-trained convolutional neural network (CNN) model to forecast 30-day temperature anomalies with 1- to 10-day lead times. The model is further fine-tuned using outputs from two sub-seasonal to seasonal (S2S) dynamic models—NCEP-CFSv2 and CMA-CPSv3—incorporating atmospheric circulation and OLR predictors. The results demonstrate that: 1) The CNN model achieves spatial anomaly correlation coefficient (ACC) exceeding 0.35 for 30-day temperature anomaly forecasts at lead times of 1 to 10 days, substantially outperforming both S2S models, which exhibit ACCs below 0.2. In addition to ACC, the CNN model consistently outperforms the dynamical models in terms of temporal correlation coefficient (TCC), anomaly sign consistency (AS), and root mean square error (RMSE). 2) Seasonal evaluation shows that the CNN model maintains superior ACC skill across all months. The AS skill is particularly higher during autumn and winter, while predictive performance declines during the winter-spring transition, early summer, and the Meiyu season. Spatially, the CNN model underperforms NCEP-CFSv2 in eastern and northwestern Hunan but performs better in other regions. At a 10-day lead time, the CNN model achieves higher TCC scores than CMA-CPSv3 at all stations across Hunan. 3) Interpretability analysis indicates that the CNN model places strong emphasis on predictors from the tropical Indian Ocean, suggesting that this region may serve as a key source of predictability. Conversely, mid- and high-latitude atmospheric circulation—which is known to influence subseasonal temperature variation in Hunan—is underutilized by the CNN model. This may partially explain reduced model performance in spring and summer. Future work should focus on enhancing model performance and robustness by incorporating additional predictors, utilizing higher-resolution reanalysis and model data, and employing more advanced deep learning architectures.

Abstract:Geostationary meteorological satellites can detect precursor signals of cumulus cloud development earlier than weather radar systems,making them valuable for convective initiation forecasting.To leverage this advantage,various algorithms have been developed,typically involving cloud detection,the removal of cirrus and mature clouds,overlap tracking,and convective initiation identification.Among these steps,the removal of cirrus and mature clouds is particularly crucial,as these cloud types can obscure developing cumulus clouds.However,existing methods face challenges such as cumulus cloud fragmentation after cirrus and mature cloud removal,difficulties in applying overlap tracking to complex cloud imagery,and limitations in threshold-based convective initiation identification.To address these issues,this study introduces several targeted improvements.First,a novel approach is proposed that treats complete cloud clusters as the primary research subject,allowing for the comprehensive extraction of cumulus lifecycle samples.Second,the Hungarian algorithm is incorporated to enhance multi-target tracking capabilities.Third,a random forest algorithm is employed to improve the accuracy of convective initiation identification.This study utilizes data from Himawari-8/9 satellites and weather radar observations to analyze convective initiation in the Shijiazhuang region.A cumulus cloud identification method,specifically tailored to the region,was developed and combined with a multi-target tracking algorithm to construct a detailed dataset of convective cells.By integrating this dataset with radar observations,cumulus clouds associated with weather processes exhibiting reflectivity values above 35 dBZ were identified.The time at which reflectivity first reached 35 dBZ was recorded as the convective initiation time,providing a robust dataset for further analysis.A comparative analysis of multi-channel brightness temperature variations and cumulus cloud development processes revealed key trends.Specifically,as cumulus clouds evolved into strong convective systems,the 10.4 μm brightness temperature in the Shijiazhuang region exhibited a decreasing trend,while the brightness temperature difference between 12.4 μm and 10.4 μm,as well as the three-channel brightness temperature difference (TTD),showed an increasing trend.These patterns were used to identify key factors influencing convective initiation.Based on these findings,a random forest model was developed for convective initiation forecasting in the Shijiazhuang region.The model demonstrated strong performance during testing,achieving a 92％ probability of detection (POD) and a 31％ false alarm rate (FAR).These results indicate that the model effectively identifies cumulus clouds likely to develop into strong convective systems,even before radar-detectable echoes emerge.A key contribution of this study is its potential to improve the timeliness of severe convective weather warnings in the Shijiazhuang region.By leveraging satellite data and advanced machine learning techniques,the proposed algorithm can detect developing cumulus clouds earlier than traditional radar-based methods.This capability is particularly valuable in regions where severe convective weather significantly impacts agriculture,transportation,and public safety.The integration of Himawari-8/9 satellite data with weather radar observations enhances the understanding of convective processes,leading to more accurate and timely forecasts.In conclusion,this study represents a significant advancement in convective initiation forecasting by addressing key challenges in cloud detection,tracking,and identification while integrating machine learning techniques.The successful application of this model in the Shijiazhuang region demonstrates its potential for broader use in other convective weather-prone areas.Future research could focus on refining the model,expanding the dataset,and exploring additional machine learning approaches to further enhance forecasting accuracy and reliability.This study not only advances the scientific understanding of convective processes but also has practical implications for improving weather warning systems and mitigating severe weather impacts.

Abstract:High-accuracy and high-resolution air temperature forecasts are essential for both theoretical research and practical applications. Mountainous terrain presents significant challenges due to its complex topography and pronounced spatial heterogeneity in air temperature distribution. Conventional numerical weather prediction models, including global and regional forecast systems, is typically operate at a spatial resolution of kilometer scale, limiting their ability to capture local-scale air temperature variations in complex mountainous terrains. Therefore, downscaling numerical model outputs is crucial for enhancing the spatial resolution and accuracy of air temperature forecasts in such areas. This study introduces the DOWN+BC method, a novel approach for downscale and bias-correcting air temperature forecasts generated by the Weather Research and Forecasting (WRF) model. This method integrates random forest downscaling, first-order adaptive Kalman filtering, and Extreme Gradient Boosting (XGBoost) models to produce high resolution (30 m) hourly air temperature forecasts with a lead time of up to 24 hours. The downscaling process begins with training a Random Forest model using WRF-forecasted air temperature as the dependent variable and 1 km resolution land surface parameters as independent variables. This trained model is then used to downscale the hourly WRF-forecasted air temperatures to a finer 30 m resolution. Finally, a first-order adaptive Kalman filter model is applied for bias correction, where the key parameters w and n are optimized through calibration. Results indicate that the DOWN+BC method effectively enhances the spatial resolution and accuracy of forecasted air temperatures in mountainous regions. The Random Forest model captures fine-scale spatial distribution patterns of near-surface air temperature more accurately, while the subsequent bias correction aligns the forecasts more closely with the actual terrain and underlying surface characteristics. Compared to the WRF forecasts, the root mean square error (RMSE) and mean absolute error (MAE) of the DOWN+BC-corrected air temperature forecasts at AWS locations decreased by 1.39 ℃ and 1.13 ℃, respectively. Additionally, in comparison with the air temperature distribution estimated by the XGBoost model, the DOWN+BC method achieved a spatial RMSE and MAE reduction of 1.19 ℃ and 0.97 ℃, respectively. The method also accurately forecasts air temperature inversions that typically occur during nighttime. Overall, the DOWN+BC method, which combines machine learning and adaptive Kalman filtering, significantly improves the spatial resolution and accuracy of WRF model forecasts in mountainous terrain. Moreover, its relatively simple implementation makes it adaptable to other regions. However, its predictive performance may be affected by abrupt and extreme temperature fluctuations, which could lead to a decrease in forecast accuracy to a certain extent.

Abstract:Numerical model predictions often contain biases when compared to local observations.Correcting these biases is essential for improving forecast accuracy.This study uses CMA-MESO model data from 2020 to 2021 (including hourly 2 m temperature,10 m wind components,sea level pressure,etc.) and observations from 340 assessment stations in Gansu Province to develop two time-lag correction models for 2 m air temperature.The correction models are based on machine learning methods,specifically LightGBM and XGBoost.The evaluation and performance analysis revealed the following:1) The accuracy of the LightGBM and XGBoost models in predicting hourly 2 m air temperature was 74.57％ and 74.33％,respectively.Both models improved prediction accuracy by 27.6％ and 27.2％ compared to SCMOC and by 53.5％ and 53.0％ compared to the CMA-MESO model.The LightGBM model slightly outperformed XGBoost,particularly in regions where CMA-MESO model performed poorly,with improvement rates exceeding 45％.2) Both models significantly reduced the forecast bias in diurnal variation for hourly 2 m air temperature compared to CMA-MESO,though performance at 07:00 BST and 16:00 BST was less accurate than at other times.3) The CMA-MESO model’s forecasted 2 m temperatures showed divergent and asymmetric distributions,while the LightGBM and XGBoost models reduced systematic biases,achieving more symmetric and convergent distributions.Both correction models decreased the number of stations with large forecast errors,and their root mean square error distribution approached an unbiased state.The LightGBM,in particular,excelled in correcting areas with significant forecast errors and in predicting temperature peaks.These results demonstrate that machine learning methods offer great potential for improving 2 m temperature predictions in numerical forecast products.

Abstract:Extended-range forecasting (i.e.,weather predictions with lead times of 10—30 days) remains a major challenge in meteorology.This study utilizes daily precipitation data from 97 meteorological stations in Hunan Province from 2005 to 2022,along with subseasonal-to-seasonal (S2S) forecast products from the European Centre for Medium-Range Weather Forecasts (ECMWF) and the National Centers for Environmental Prediction (NCEP).The dataset is divided into a training and validation period (2005—2018) and an independent prediction period (2019—2022) to investigate extended-range precipitation forecasting.To address the systematic biases inherent in numerical models for extended-range forecasting,this study proposes a two-stage correction scheme.In the first stage,a hierarchical cumulative probability matching method is employed to adjust the forecast distribution and reduce systematic errors.In the second stage,a low-frequency threshold method is applied to further refine the forecasts,particularly for extreme precipitation events.Building upon this bias-corrected dataset,a convolutional neural network (CNN) model is developed to capture the coupling relationships between large-scale circulation patterns and precipitation fields.The CNN models are trained separately using the ECMWF and NCEP forecast products.Experimental results show that bias correction significantly improves the accuracy of extended-range precipitation forecasts for both models,with the NCEP model exhibiting greater improvement.Specifically,the threat score (TS) for single-station precipitation forecasts from the NCEP model increases by 38.5％,while its regional precipitation score improves by 43.9％.In comparison,the TS for the ECMWF model improves by 14.0％,and its regional precipitation score increases by 24.2％.Independent predictions indicate that the ECMWF model generally outperforms the NCEP model,particularly before for lead times of less than 15 days,although the performance gap narrows as the lead time increases.Furthermore,CNN-based models demonstrate superior forecasting skill compared to standalone numerical models,particularly for lead times of 15—30 days,where their advantages become more pronounced.In long-term extended-range precipitation forecasting,the CNN model,trained on historical data,effectively identifies precursor signals of precipitation,thereby significantly improving forecast accuracy.To further optimize forecasting performance,an integrated prediction approach is proposed.This method combines the corrected numerical model forecasts with the CNN-based forecasts using a weighted integration technique,balancing the strengths of each model while minimizing errors.The integrated forecasts exhibit high predictive skill across the entire 10—30 days extended-range period,providing a reliable basis for operational precipitation forecasting.Future research will focus on extending the model to regions with diverse climatic conditions,exploring more advanced machine learning techniques,and incorporating additional predictive factors closely related to precipitation.These efforts aim to enhance the model’s generalizability and forecasting precision.

Abstract:This study introduces a residual spatio-temporal stacking (Res-STS) model designed to improve temperature forecasting in Hunan Province,China—a region characterized by complex terrain,with mountainous areas to the east,west,and south,and the Dongting Lake Plain to the north.This diverse topography,influenced by elevation gradients,vegetation cover,cold air pooling,and lake effects,results in spatially heterogeneous and temporal dynamic temperature patterns.
Although deep learning models such as ResNets have demonstrated success in precipitation forecasting and severe weather recognition,their application to temperature forecasting remains limited—particularly regarding the integration of multi-scale physical variables from numerical models.To address this gap,the Res-STS model integrates both surface and upper-air variables from the European Centre for Medium-Range Weather Forecasts Integrated Forecasting System (ECMWF-IFS),along with observational data,thereby enhancing the model’s ability to capture spatiotemporal dependencies.
The Res-STS architecture adapts the ResNet with residual connections to mitigate gradient vanishing and explosion,preserving shallow-layer features.Unlike conventional sequential temporal models,Res-STS employs a spatiotemporal stacking approach to jointly learn background environmental fields and temporal evolution patterns.Spatially,a “field-to-point” framework is adopted:a 250 km×250 km region centered on each forecast grid point—corresponding approximately to meso-beta-scale systems—is used as input.This design balances computational efficiency and the retention of large-scale atmospheric information,avoiding the limitations of point-to-point oversimplification and field-to-field data scarcity.Temporally,consecutive 3-hourly ECMWF-IFS forecast fields are stacked to predict hourly temperatures in sequential time windows (e.g.,forecasts at T0 and T1 are used to predict hours 1—3;T1 and T2 for hours 4—6,and so on),enabling the generation of continuous 24-hour forecasts.
Evaluation results show that the Res-STS model outperforms benchmark models across all tested metrics.Compared with ECMWF-IFS and guidance from the National Meteorological Centre,Res-STS achieves mean absolute errors (MAEs) of 1.21 ℃ for hourly forecasts,1.38 ℃ for daily maximum temperatures,and 1.07 ℃ for daily minimum temperatures—representing reductions of 23.8％ and 15.2％ in MAEs for maximum and minimum temperatures,respectively.In high-altitude areas above 800 meters,the model’s median error (1.12 ℃) is 31％ lower than that of ECMWF-IFS.During extreme events,such as cold waves and heatwaves,Res-STS also outperforms manual corrections and objective forecasts,achieving 2 ℃ accuracy rates of 85.81％ for minimum temperatures and 97.88％ for maximum temperatures.
Nonetheless,the model's reliance on ECMWF-IFS input fields constrains its performance under systematic biases—such as errors in cloud cover estimation—which can increase MAEs by 0.8—2.3 ℃ during persistent synoptic anomalies.Additionally,the 0.05° resolution of the China Meteorological Administration Land Data Assimilation System (CLDAS) dataset may smooth terrain transitions in narrow valleys,contributing to residual afternoon temperature discrepancies of up to 1.8 ℃.Current computational limitations restrict the model's operational use to Hunan Province.Future research is needed to reduce reliance on numerical model inputs,incorporate higher-resolution terrain data,and optimize computational performance for broader deployment and improved forecast accuracy during extreme cooling events.
This study advances the integration of deep learning with numerical weather prediction and offers a novel post-processing framework for temperature forecasting in regions with complex topography.

Abstract:In the context of accelerating global climate change and increasingly frequent extreme weather events, accurate weather and climate prediction has become critically important for safeguarding human society and the natural environment.Traditional meteorological models face significant challenges in simulating complex atmospheric systems and handling high-dimensional data, limiting the accuracy and reliability of both short-term weather forecasts and long-term climate projections.In recent years, revolutionary advancements in artificial intelligence (AI), particularly in generative models, have shown remarkable potential across a range of scientific domains.Among these, diffusion models have emerged as a particularly promising approach, owing to their stable training processes, unique generative mechanisms, and superior sample quality.This paper provides a comprehensive review of diffusion models and their applications within meteorological science, focusing on their current implementation, performance characteristics, and future prospects across multiple meteorological subfields.
The review first establishes the theoretical foundations of diffusion models, detailing their core mechanisms, including the forward and reverse diffusion processes.It then explores the intrinsic connections between diffusion models and meteorological science, emphasizing shared mathematical frameworks and physical principles such as stochastic processes, Bayesian inference, and multi-scale dynamics.Building on this foundation, the paper systematically examines four key application areas where diffusion models have demonstrated particular promise.In precipitation forecasting, models such as LDCast, PreDiff, GED, and DiffCast have achieved significant improvements in generating diverse precipitation scenarios and accurately capturing extreme events.For data assimilation, approaches such as SDA, DiffDA, and SLAMS integrate heterogeneous, sparse, and noisy observational data into weather fields while significantly reducing computational costs compared to traditional methods.In spatial downscaling, models like CorrDiff, StormCast, and STVD enhance spatial resolution from coarse global climate models to fine-scale regional predictions while maintaining physical consistency and capturing local terrain effects.In weather system simulation, models such as SEEDS and GenCast have shown strong capabilities in modeling complex atmospheric dynamics and producing high-quality ensemble forecasts that reflect the inherent uncertainties in weather prediction.
Despite these advancements, several challenges remain, including high computational demands, maintaining physical consistency, ensuring long-term model stability, and improving model interpretability to foster greater trust among meteorologists.Looking ahead, the paper identifies five promising research directions:climate change scenario generation, multi-scale weather-climate joint modeling, meteorological data completion and reconstruction, linking global and regional scales, and uncertainty quantification in meteorological modeling.These emerging applications span a wide range of temporal and spatial scales, from microscale to macroscale and from short-term forecasts to long-term projections.The integration of diffusion models with traditional physical physics-based weather models holds significant potential for improving forecast accuracy, enhancing climate model resolution, and advancing extreme weather early-warning systems.As technology innovation progresses and interdisciplinary collaboration deepens, diffusion models are expected to play an increasingly critical role in meteorological research and operational applications, offering powerful new tools for understanding atmospheric phenomena and addressing the challenges of climate change.