Abstract This study evaluates Pangu-Weather, a machine-learning-based weather prediction (MLWP) model, for its ability to reduce errors in numerical predictions of significant wave height (SWH). SWH predictions from the SWAN wave model were driven by 10 m wind fields from Pangu-Weather, and error assessments were conducted at 50 observation stations along the Japanese coast over 52 seven-day periods in 2022. For comparison, meteorological forcing from the Weather Research and Forecasting (WRF) numerical weather prediction (NWP) model was also applied. When used as meteorological forcing, Pangu-Weather performed comparably to WRF in most cases, but showed a clear advantage for observed SWH < 1.0 m. In contrast, WRF showed better performance for observed SWH ≥ 1.0 m. This is partly because Pangu-Weather often struggles to predict severe meteorological events (e.g., typhoons and bomb cyclones) and their associated SWH. These findings, within the scope of this study, indicate that MLWP forcing has the potential to serve as an alternative to NWP for SWH prediction, while further work is needed to extend this skill to rarer high-impact events.

The potential of AI global weather models for reference evapotranspiration forecasting: a comparison with numerical weather prediction models

Accurate wind speed forecasts are critical for integrating wind energy into power grids, reducing imbalance costs in electricity markets, and optimizing wind farm operations. Day-ahead forecasts are typically generated using numerical weather prediction (NWP) models. This work proposes a hybrid model for 24-hour wind speed forecasting, which combines the Weather Research and Forecasting (WRF) model with the Seasonal Autoregressive Integrated Moving Average (SARIMA) model. The proposed model improves the accuracy of the WRF wind speed forecast through the SARIMA technique by identifying significant autocorrelations in the forecast errors. The study was conducted in La Ventosa, Mexico, a region with significant development in the wind power sector. Wind speed data measured at heights of 17.5 m and 40 m were used during periods of low and high wind speeds. The model’s performance was evaluated using the metrics mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE). The results showed that the hybrid WRF-SARIMA model outperformed the WRF model. Forecast errors for MAE were reduced between 29% and 45%, for MSE between 40% and 67%, and for RSME between 22% and 43%. The WRF-SARIMA leverages the benefits of physical NWP models while incorporating the interpretability and reduced computational cost of traditional statistical models. In this way, the proposed model improves wind speed forecast accuracy, especially in the operational contexts of wind energy management.

Abstract Sea fog poses a significant risk to marine transportation by reducing visibility and affecting navigation efficiency. Traditional methods for predicting sea fog face many challenges due to the complex nature of fog formation and the inherent limitations of current numerical and statistical models. To improve the accuracy of sea fog prediction, we present a hybrid forecasting approach that combines a machine learning (ML) model with the numerical weather prediction (NWP) model. We introduced the time-lagged correlation analysis (TCA) to identify critical predictors and mechanisms of fog formation. To address data imbalance issues, we applied the focal loss function and ensemble learning. A hybrid forecasting model was developed for the Yangtze River Estuary (YRE) and is suitable for various locations and seasons. It can predict sea fog events with visibility below 1 km up to 60 h in advance. Our method outperforms traditional statistical approaches, achieving notable improvements in the equitable threat score (ETS) and the Heidke skill score (HSS). This study demonstrates the potential of ML to improve the accuracy of sea fog forecasts. Significance Statement This study aims to enhance sea fog prediction, a hazard for marine navigation due to reduced visibility. Traditional methods struggle with accurate forecasts. We introduce a hybrid approach combining a machine learning model with a numerical weather prediction model, predicting sea fog with visibility under 1 km up to 60 h ahead. Our method surpasses two traditional statistical approaches, showcasing machine learning’s (ML’s) potential to boost forecast accuracy.

Abstract Determining surface precipitation type is an operational forecasting challenge, particularly in complex cold-season precipitation events and in areas with a dearth of observations. This paper describes a new radar algorithm proposed for the NEXRAD network called the surface hydrometeor classification algorithm (sfcHCA). The sfcHCA produces estimates of surface precipitation types within the radar surveillance area by merging output from the existing NEXRAD HCA level-III product with surface precipitation type diagnosed from a 1D steady-state discrete particle model initialized from numerical weather prediction model data. The sfcHCA introduces three new HCA classes for freezing rain, ice pellets, and freezing rain/ice pellet mixtures, as well as several novel features such as the utilization of range-defined quasi-vertical profiles, polarimetric radar particle size distribution retrievals for model initialization, merging logic for combining model-diagnosed surface precipitation-type classifications with existing HCA output, and a radar-melting-layer-based adjustment algorithm to correct for erroneous background model data. The sfcHCA was validated against more than 15 000 manned ASOS station reports and over 30 000 Mobile Precipitation Identification Near the Ground (mPING) reports across 225 cases spanning all CONUS NEXRAD sites. Probability of detection rates within 20 km of manned ASOS observations were 97.1%, 78.3%, 72.7%, 68.1%, and 56.0% for rain, snow, rain/snow mix, freezing rain, and ice pellets, respectively. This algorithm should provide useful information for nowcasting surface-precipitation-type evolution and offering impact-based decision support. Significance Statement Awareness of surface precipitation type is a critical need for operational forecasters. Weather radars are a well-suited tool for observing precipitation aloft but must be combined with other environmental data to infer precipitation type below the radar beam. This study proposes a new product for the NEXRAD network that combines radar data with environmental modeling to produce estimates of surface precipitation type throughout the radar viewing area, including for high-impact and hard-to-predict precipitation types such as freezing rain and ice pellets. This product will provide additional situational awareness for forecasters during complex and evolving precipitation events, particularly in regions with sparse observations.

Climate change poses both pressing scientific and societal challenges requiring accurateprediction of extreme events and rigorous adaptation approaches. While classical methodssuch as numerical weather prediction (NWP) and general circulation (GCM) models are stillcentral to climate depictions, these models are computationally expensive and, therefore,struggle at real-time applications and inherently, in their accuracy .New advancements in artificial intelligence (AI) suggests that, at the very least, machinelearning (ML) and deep learning (DL) could provide a revolutionary and complementaryalternative to the physics-based modelling shown previously. In this review, we outline 70peer-reviewed papers (2019-2025), we selected the studies using the literature reviewaccording PRISMA to cover the AI application in prediction. Which highlight the risk ofusing AI for climate forecasting and adaptation. Quantitative evidence indicates thatGraphCast surpasses ECMWF HRES in roughly 90% of forecasting metrics; GenCastdelivers 97% higher accuracy compared with ensemble means; and MetNet-2 extendsprecipitation forecasting horizons to nine hours with improved precision. In adaptation, AIhas helped predict agricultural yield with up to 88% accuracy, alert farmers of imminentdrought two months in advance, provide early warnings of dengue fever with a reported AUCof 0.89, and improve urban flood resilience with accuracy levels as high as 92% . This reviewfocus on both chances and challenges and highlight the constraints in AI applications like thedataset and difficulty in weather explanation models.However, despite these advances, challenges remain due to data bias, limited visibility intodeep learning models, high energy consumption, and unequal access to technology. Thereview sets out an evaluation of opportunities and challenges from which AI might beeffectively applied to climate science and climate-related policy, and suggests a three-pillarframework incorporating sustainability, transparency and equity which can enhanceresponsible use within climate adaptation and resilience

ABSTRACT We present the CIENS dataset, which contains ensemble weather forecasts from the operational convection‐permitting numerical weather prediction model of the German Weather Service. It comprises forecasts for 55 meteorological variables mapped to the locations of synoptic stations, as well as additional spatially aggregated forecasts from surrounding grid points, available for a subset of these variables. Forecasts are available at hourly lead times from 0 to 21 h for two daily model runs initialised at 00 and 12 UTC, covering the period from December 2010 to June 2023. Additionally, the dataset provides station observations for six key variables at 170 locations across Germany: pressure, temperature, hourly precipitation accumulation, wind speed, wind direction, and wind gusts. Since the forecasts are mapped to the observed locations, the data is delivered in a convenient format for analysis. The CIENS dataset complements the growing collection of benchmark datasets for weather and climate modelling. A key distinguishing feature is its long temporal extent, which encompasses multiple updates to the underlying numerical weather prediction model and thus supports investigations into how forecasting methods can account for such changes. In addition to detailing the design and contents of the CIENS dataset, we outline potential applications in ensemble post‐processing, forecast verification, and related research areas. A use case focused on ensemble post‐processing illustrates the benefits of incorporating the rich set of available model predictors into machine learning‐based forecasting models.

Weather forecasting plays a crucial role in numerous sectors, ranging from agriculture and transportation to disaster management. Traditional meteorological models, while valuable, often face challenges in accurately predicting complex and dynamic weather patterns. This research paper explores the integration of machine learning (ML) techniques into weather forecasting to enhance predictive accuracy and reliability. The study begins by providing an overview of the limitations of conventional numerical weather prediction models and emphasizes the need for innovative approaches. It introduces a weather forecasting system based on machine learning, utilizing Decision Tree, Support Vector Machine, Random Forest, K-Nearest Neighbors, Gradient Boosting, Logistic Regression, and Naïve Bayes algorithms. The paper discusses the development and training of ML models using large datasets to capture intricate relationships among atmospheric variables. Among the evaluated models, Gradient Boosting achieves the highest predictive accuracy by effectively capturing nonlinear relationships and minimizing prediction errors. Performance evaluation demonstrates that integrating multiple machine learning techniques provides a stable, reliable, and scalable solution for short- to medium-term weather forecasting.

This study proposes a fully automated method for typhoon area identification from numerical weather prediction (NWP) model data using the RetinaNet model combined with the ResNet152 backbone network. The input training data consisting of physical fields such as pressure, wind, and temperature from the reanalysis are processed into multi-channel images to train an object detection model in which the data labels are determined from the Japan Meteorological Agency (JMA) best-track standard typhoon trajectory data. Experimental results on reanalysis data show that the model achieves high accuracy (detection rate over 90%), thereby identifying well the information areas related to storm activity without the need for initial information (first guess) like the traditional tracking method. When applied to climate change simulation data, the model still effectively detects cyclones, reflecting the trend of spatial-temporal changes in storm activity in the future. The experiment in the Bien Dong Sea region showed realistic detection capabilities, while also identifying some weak systems that were missed by traditional algorithms. The results show that deep learning has great potential to automate storm detection from NWP data without the need for subjective initial processing.

Short-term weather prediction plays a critical role in supporting decision-making across sectors such as agriculture, transportation, and disaster risk management. This study proposes an interpretable and computationally efficient weather forecasting approach based on linear system modeling combined with Singular Value Decomposition (SVD). Historical meteorological data—including temperature, humidity, air pressure, and wind speed—are represented in matrix form to extract dominant patterns and construct a system of linear equations describing inter-variable relationships. The resulting model is evaluated for short-term forecasting horizons of 24–48 hours using standard performance metrics. Experimental results demonstrate that the proposed SVD-based linear system model outperforms conventional linear regression, achieving lower MAE and RMSE values and higher coefficients of determination (R² = 0.94 for temperature and 0.91 for humidity). While not intended to replace physics-based numerical weather prediction models for long-term forecasting, the proposed approach offers significant advantages in computational speed, interpretability, and applicability in data- and resource-constrained environments. These findings indicate that matrix-based linear system analysis provides a viable alternative for fast and accurate short-term weather prediction and can be further enhanced through integration with non-linear or machine learning-based methods.

Abstract Forecasting the location and intensity of strong winds associated with midlatitude cyclones is important as they can have significant safety, economic, and environmental impacts. In this study we use a feature-based evaluation method to assess the performance of both numerical weather prediction and machine learning weather prediction (MLWP) models in forecasting midlatitude cyclone winds. By tracking over 1000 cyclones across the Northern Hemisphere from 1 October 2023 to 31 March 2024 in 7 MLWP models, we systematically compare model performance. Our results show that MLWP models predict midlatitude cyclone tracks with accuracy comparable to the ECMWF IFS-forecast out to 10 days. However, MLWP models exhibit a persistent intensity bias, underestimating cyclone minimum sea level pressure by more than 5 hPa at 10-day forecast lead times, whereas the ECMWF IFS-forecast has no bias. Additionally, all MLWP models produce weaker than observed peak 10-m winds, even at short lead times. In contrast, the ECMWF IFS forecast exhibits no bias in 10-m wind speed. These differences highlight the limitations of current MLWP models in capturing important high-impact weather features like peak wind speeds.

Weather forecasting is a crucial point of discussion in agriculture, aviation, disaster management, and energy decision-making. Conventional numerical weather prediction (NWP) models are efficient but need a lot of computer resources and can be inaccurate in the beginning. Machine learning (ML) has become a powerful alternative and addition to traditional forecasting approaches, giving data-driven methodologies that can simulate complex meteorological patterns more efficiently and adaptably. This review covers weather forecasting machine learning methods, including supervised learning models like Support Vector Machines (SVM), Random Forests, and Gradient Boosting and deep learning frameworks like CNNs, RNNs, and LSTM networks. This review focuses on using these models to anticipate temperature, precipitation, wind speed, and humidity across various temporal and spatial scales. Modern model designs, hybrid models, ensemble learning, and satellite and sensor data to improve forecast accuracy are examined. It addresses crucial issues such data quality and accessibility, model interpretability, overfitting, and real-time forecasting. Explainable AI and uncertainty quantification are crucial to trusting machine learning-based weather systems, according to the review. This paper reviews the current and future state of machine learning in meteorological forecasting.

Abstract. Snow avalanches are a serious threat to traffic in the northern Gaspésie region. In this study, we look at the development of different forecasting models using machine learning (ML), based on snow avalanche events recorded by Québec's Ministry of Transportation, meteorological data from the Cap-Madeleine station and Environment Canada weather forecast data. The models were trained and tested on Train and Test datasets with meteorological and weather forecasts recorded at the Meteorological Station. Unsupervised learning models were compared to expert models where only 4 variables were selected with avalanche expertise in mind, yielding similar results in prediction. The ML models were then tested in a realistic forecasting context over the year 2019 with weather data from a forecasting station (Hindcast) and with weather forecast data over 24 and 48 h. The LR and RF models show that model performance can match or exceed that of current forecasting tools, enhancing hazard anticipation while maintaining a user-friendly framework suitable for real-time application. In conclusion, recommendations on forecast-based operational procedures are proposed.

Accurate hyper local weather forecasting remains one of the most challenging tasks in meteorology due to high spatial variability, changing climate patterns, and limitations of conventional statistical approaches. Traditional forecasting systems largely rely on satellite observations, numerical weather prediction models, and coarse-grained climate datasets, which fail to provide localized predictions required for agriculture, logistics, disaster management, and urban planning. Green Sentry is an AI-powered hyper local weather prediction and alert system designed to overcome these limitations by leveraging machine learning, spatiotemporal analytics, and climate datasets. The system predicts precipitation, classifies weather risks, and triggers alerts in real time. Green Sentry integrates LSTM-based time-series modeling, ensemble machine learning for anomaly detection, GIS-based visual mapping, and a web interface that delivers localized in sights at street-level granularity. Evaluated using datasets from Open-Meteo API, World Bank CCKP, and regional weather logs, the system achieves an MAE of 3.1mm, an RMSE of 5.2 mm, and a heavy rainfall prediction accuracy of 90.1%. Experimental results demonstrate improvements in prediction precision, response time, and contextual climate awareness compared to baseline models. The proposed system supports decision-making in agriculture, flood management, irrigation planning, transportation, and community safety. This paper presents the system architecture, model implementation, evaluation, limitations, and future enhancements toward climate-resilient, AI-driven local forecasting tools.

The accuracy and timeliness of precipitation inputs have significant impact on flood forecasting. Upstream Minjiang River Basin is characterized by complex terrain and highly variable climatic conditions, posing a significant challenge for runoff forecasting. This study proposes a combined forecasting approach integrating numerical weather prediction (NWP) models with hydrodynamic models to enhance flood process simulation. The most appropriate initial field data for the Weather Research and Forecasting Model (WRF) exist in time and space resolution. Compared with the measured series, the characteristics of precipitation forecasting are summarized from practical and scientific perspectives. InfoWorks ICM is then used to implement runoff generation calculations and flooding processes. The results indicate that the WRF model effectively simulates the spatial distribution and peak timing of precipitation in the upper Minjiang River. The model systematically underestimates both peak rainfall intensity and cumulative precipitation compared to observations. Initial field data with 0.25° spatial resolution and 3 h temporal intervals demonstrate good performance and the 10–14 h forecast period exhibits superior predictive capability in numerical simulations. Updates to elevation and land use conditions yield increased cumulative rainfall estimates, though simulated peaks remain lower than measured values. The runoff results could indicate peak flow but rely on the precipitation inputs.

Abstract Accurate precipitation forecasting often relies on high-resolution numerical weather prediction (NWP) models, which are essential for capturing fine-scale and nonlinear atmospheric dynamics. However, the computational demands of these models can be substantial. Leveraging recent advancements in artificial intelligence (AI), we present a stretched-grid AI-driven weather model with 6-km horizontal grid increments over the Western United States and ~31 km in other regions globally. The model employs an autoregressive framework to generate forecasts in minutes and is evaluated against global and regional NWP systems, as well as a lower-resolution AI model. Our results show that the regional AI model reduces 24-h accumulated precipitation errors, performs competitively with the regional NWP model, and effectively captures extreme precipitation events, particularly those linked to atmospheric rivers, which global coarser models often underestimate. This work underscores the potential of regional, high-resolution AI models for precipitation forecasting at km-scales, and discusses some of the challenges for future development.

Abstract. Contrails formed by aircraft in ice supersaturated regions (ISSR) can persist and spread for several hours, evolving into cirrus which have a net positive effect on global warming. Reducing this contribution could be achieved through on-purpose flight planning, in particular by avoiding ice supersaturated regions. In this context, a modification to the cloud scheme of the ARPEGE (Action de Recherche Petite Echelle Grande Echelle) operational numerical weather prediction (NWP) model is proposed to enable the representation of ISSRs at cruise altitude. This modification does not require any major algorithmic changes or additional computational effort, and the methodology is transferable to similar parameterizations, commonly used in global circulation models. Humidity forecasts are evaluated using in situ aircraft humidity observations and compared with operational forecasts from ARPEGE and the Integrated Forecast System (IFS). A sensitivity study on neighborhood tolerance and humidity thresholding is carried out, enabling a comprehensive comparison between NWP forecasts. It is shown that the modified cloud scheme allows for supersaturation, significantly improving the representation of humidity with respect to ice, with ISSR discrimination skills close to IFS (hit rate ∼ 80 % and false alarm ratio ∼ 30 % when a neighborhood tolerance of 150 km, i.e. 10 min of flight, is applied). The spatial correspondence between observations and the modified ARPEGE model is illustrated by a commercial flight case study. The modelization of ice supersaturation in ARPEGE can therefore be used for further contrail climate impact applications, together with the associated evaluation methodology, which contributes to the definition of a shared framework for ISSR verification.

Weather forecasting is a core task in meteorology with significant impacts on socio-economic activities. While traditional numerical weather prediction models are foundational, they are computationally intensive and time-consuming. In response, data-driven approaches are gaining prominence for their efficiency and accuracy. This study proposes DDCT, a novel data-driven autoregressive forecasting model that synergistically integrates Convolutional Neural Network (CNN) and Transformer architectures. Central to our design is a parallel architecture where a CNN module captures fine-grained local spatial features, while a Transformer module simultaneously models global spatio-temporal dependencies. Leveraging an autoregressive mechanism, DDCT extends the prediction horizon through step-by-step forecasting, making it highly effective for longer-term weather prediction. We validate the model’s effectiveness through extensive experiments on the ERA5 and SEVIR meteorological datasets. The results demonstrate that our method significantly outperforms existing baseline models in predicting multiple weather variables, especially for challenging long-term forecasts. This research not only provides new insights for hybrid deep learning in meteorology but also establishes DDCT as a robust framework with significant application value and research potential.

This paper reports on the study of turbulence at various locations in Hong Kong during Typhoon Wipha in July 2025, including turbulence intensity based on Doppler Light Detection and Ranging (LIDAR) systems and radiosondes, observations by microclimate stations, and low-level windshear and turbulence at the Hong Kong International Airport (HKIA) by LIDAR, flight data, and pilot reports. Although the observation period was primarily limited to 20 July 2025, passage of a typhoon over a densely instrumented urban area is uncommon; these observations on turbulent flow associated with typhoons therefore can serve as valuable benchmarks for similar studies on turbulent flow associated with typhoons in other coastal areas, particularly for operational alerts in aviation. To assess the predictability of turbulence, the eddy dissipation rate (EDR) was derived from a high-resolution numerical weather prediction (NWP) model using diagnostic and reconstruction approaches. Compared with radiosonde data, both approaches performed similarly in the shear-dominated low-level atmosphere, while the diagnostic approach outperformed when buoyancy became important. This result highlights the importance of incorporating buoyancy effects in the reconstruction approach if the EDR diagnostic is not available. The high-resolution NWP was also used to provide time-varying boundary conditions for computational fluid dynamics simulations in urban areas, and its limitations were discussed. This study also demonstrated the difficulty of capturing low-level windshear encountered by departing aircraft in an operational environment and demonstrated that a trajectory-aware method for deriving headwind could align more closely with onboard measurements than the standard fixed-path product.

Abstract Simultaneous lidar and radar observations of clouds and precipitation are becoming more common, and to take advantage of such opportunities, radar–lidar-estimated diameter (RLED) was introduced in earlier work to describe relevant droplet size. Here, we use in situ cloud and drizzle probe measurements of drop size distribution (DSD) to show that RLED, the ratio of the sixth to the second moment of DSD ( D 62 ), is sensitive to collision–coalescence, making it a potential indicator of drizzle onset. Remarkably, despite differences in sample volumes, in situ estimates of RLED agree closely with those based solely on lidar–radar measurements. Additionally, RLED is sensitive to DSD dispersion as we demonstrate on remote sensing observations. This is significant as DSD dispersion plays an essential role in numerical weather prediction models and terrestrial radiative transfer. Significance Statement We used in situ measurements of water droplets and radar and lidar measurements to explore the onset of drizzle. To better understand these observations, we used radar–lidar-estimated diameter (RLED) bridging in situ and remote measurements, which helps describe the size of water droplets in clouds. RLED is a promising tool for indicating when light rain or drizzle begins, helping to detect drizzle onset.