This session seeks contributions exploring the development and application of ML or ML-enhanced algorithms for forecasting weather and climate at multiple spatial and temporal scales and for detecting and anticipating extreme weather and climate events. We welcome studies that address the use of AI for short-and medium-range meteorological forecasts, extended-range forecasts, sub-seasonal to seasonal climate forecasts, or longer-term climate projections, spanning local to global spatial scales.
We particularly encourage submissions that connect extremes to their societal and environmental impacts, such as impacts on infrastructure, ecosystems, health, or energy systems.Contributions that integrate ML with physical mechanisms to advance the representation of climate variables in numerical models or climate datasets are also highly encouraged.
By bringing together experts from AI, data science, meteorology, climate science, and impact modelling, this session aims to foster interdisciplinary collaborations that push the boundaries of forecasting and understanding extreme weather and climate events, as well as their impacts. We encourage submissions from early-career scientists, established researchers, and industry professionals alike.
Orals: Wed, 6 May, 16:15–18:00 | Room 0.31/32
Accurate subseasonal forecasting of drought indices across spatial and temporal domains in Europe remains a major challenge due to internal climate variability, the inherent uncertainty in AI-driven forecasts, and complex atmospheric interactions. These challenges are particularly pronounced for rare and severe drought events, which can have substantial societal and environmental consequences. Recent advances in machine learning have improved climate forecasting, but the contribution of internal climate variability to predictive uncertainty in drought forecasts remains insufficiently quantified.
This study investigates whether observed limitations in the predictive performance of AI-based subseasonal drought forecasts can be explained by internal climate variability. To address this, we develop a Temporal Fusion Transformer framework to forecast the Standardized Precipitation–Evapotranspiration Index for a single month (SPEI-1) over the European domain. We extract the internal variability of a regional climate model large ensemble and quantify the extent to which predictive imprecision is attributed to internal climate variability. This approach enables a systematic assessment of hot and dry extremes, forecast skill, and uncertainty characterization.
The proposed approach enhances existing forecasting methods, particularly in terms of uncertainty quantification and its effective communication. The Temporal Fusion Transformer captures key temporal and spatial characteristics of SPEI-1 variability across Europe, except for limitations over the complex terrain of the Alps. Analysis of forecast variability shows that a substantial fraction of predictive uncertainty can be attributed to internal climate variability rather than model deficiencies alone.
The interpretable uncertainty bounds provide a tool supporting risk assessment and decision-relevant drought forecasting, because they highlight the important role of internal climate variability for drought prediction. Overall, this work emphasizes how merging AI-driven forecasting techniques with quantification of internal climate variability can support more reliable and decision-relevant assessments of drought risk.
How to cite: Funk, H., Gruber, C., Kauermann, G., Küchenhoff, H., Ludwig, R., and Mittermeier, M.: Uncertainty-Aware AI Forecasting of European Droughts: The Role of Internal Climate Variability, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10474, https://doi.org/10.5194/egusphere-egu26-10474, 2026.
Recent deep learning advances improve predictive performance but often increase computational and memory costs. This limits use in resource-constrained settings. Meanwhile, meteorological data exhibit strong multiscale characteristics. Training such signals with single-scale models can cause scale mixing and spectral bias, which degrade performance in extreme events and long-term forecasting.
Motivated by these challenges, this study explores an alternative strategy that enhances forecasting performance through scale-aware data preprocessing rather than increased model complexity. Multivariate Variational Mode Decomposition (MVMD) is integrated with graph neural networks (GNNs) to separate multi-scale temporal variability before spatial learning. Surface wind forecasting over Taiwan is characterized by complex atmospheric dynamics associated with typhoons, Meiyu fronts, and monsoon systems. It provides a challenging case for 72-hour wind speed forecasting.
ERA5 reanalysis data at a 0.25° spatial resolution and 12-hourly intervals over East Asia (5–40°N, 105–140°E) are used to construct a scale-aware spatio-temporal forecasting framework. The training dataset spans 2000 to 2016, the validation dataset spans 2016 to 2020, and the testing dataset spans 2020 to 2024. Raw surface wind fields are decomposed into five intrinsic mode functions (IMFs) using MVMD, with the number of modes selected based on a balance between root mean square error (RMSE), signal-to-noise ratio (SNR), and orthogonality index (OI). These scale-separated wind components with selected background meteorological variables (temperature, mean sea-level pressure, sea surface temperature, and 500-hPa variables) are incorporated into a three-layer Graph Attention Network (GATv2) model. The model is trained for one-step-ahead prediction, and multi-day forecasts are generated through an autoregressive rollout strategy that does not rely on additional temporal sequence encoders.
The MVMD–GATv2 model was compared with a baseline GATv2 trained directly on raw surface wind fields. Model performance is evaluated using mean absolute error (MAE), RMSE, and anomaly correlation coefficient (ACC). Preliminary results show that RMSE at the 12-hour forecast point decreased from 1.7 to 0.8. In addition to improved accuracy, ongoing analyses within this comparison framework focus on examining the evolution of errors across lead times and quantifying training costs. Further analyses assess the interpretability of scale-separated representations and explore boundary-related effects. In summary, these findings highlight the potential of MVMD as a scale-aware data preprocessing strategy that improves the accuracy, stability, and interpretability of graph-based regional wind predictions.
How to cite: Cheng, J. and Tsai, C. W.: A Scale-Aware Graph Neural Network Framework via Multivariate Variational Mode Decomposition for Multi-Day Wind Speed Forecasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18249, https://doi.org/10.5194/egusphere-egu26-18249, 2026.
Recent AI weather models outperform traditional physics-based weather prediction models on many benchmarks. The evaluation is mostly restricted to point-wise metrics such as the mean squared error and therefore does not assess whether the joint multivariate behavior is well captured. Since AI weather models do not rely on any physical laws, there are strong concerns and first indications that the forecasted fields lack physical consistency in terms of spatial coherence and energy constraints. Verifying such constraints directly is however far from trivial.
We propose a Turing test for physicality that leverages the spread of an ensemble of pre-trained AI forecasting models. The main idea is that the epistemic uncertainty of these models is much larger when applied to non-physical conditions compared to physical conditions that have been part of the training data. We combine this intuition with the theory of conformal inference to obtain a statistical test for physicality with finite-sample guarantees. Case studies on the 1963 Lorenz system show the effectiveness of our proposed approach in identifying conditions that lie outside of its attractor. We then illustrate the applicability of our methodology to recent AI weather models.
How to cite: Engelke, S., Gnecco, N., Froelich, M., Hentschel, M., and Zhang, Z.: A Turing test for physicality in AI weather models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18411, https://doi.org/10.5194/egusphere-egu26-18411, 2026.
In recent decades, extreme meteorological events have increased in frequency and intensity, enhancing hydrogeological risk. This study evaluates the performance of a Machine Learning model based on a Latent Diffusion Network (Latent Diffusion Model, LDM), developed within the RETE project, a joint initiative of FBK and ENEA, in generating high-resolution precipitation fields over Italy. Four historical precipitation datasets produced by the LDM are compared with the main reanalysis products, ERA5 and CERRA, to assess their ability to reproduce precipitation climatology and extreme events. The analysis is based on standard climatological statistics and Extreme Value Theory (EVT). Climatological features are examined through daily mean and seasonal cumulative precipitation, while extremes are investigated by estimating precipitation levels associated with 10, 20, and 50-year return periods. The results provide insight into the reliability of LDM-based products as complementary tools to traditional reanalyses for climate studies and potential operational applications.
How to cite: Perilli, L., Calmanti, S., and Petitta, M.: Analysis and comparison of extreme precipitation events between physical models and Artificial Intelligence models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10341, https://doi.org/10.5194/egusphere-egu26-10341, 2026.
MAR is a Regional Climate Model (RCM) used over Belgium that provides deterministic downscaling of reanalyses and Earth System Models (ESMs) at 5-km resolution (Doutreloup et al., 2019). These high-resolution fields are computationally expensive to produce as they require solving complex physical equations. Combined with its deterministic nature, this limits the use of MAR for assessing the frequency and intensity of extreme events and their future changes.
To address this limitation, we have developed MAR.ia, a diffusion-based emulator of MAR which provides probabilistic estimates of downscaled fields at a lower computational cost (from 0.25° and 1° ERA5 fields). This allows the direct generation of ensembles from which we can derive a range of possible weather outcomes and estimate their corresponding likelihood.
However, the reproduction of extreme events is expected to be more challenging for diffusion models because these events might be scattered or absent from the training set. This is due to the fact that they are rare, and also to climate change which induces a shift between the training and testing distributions.
We evaluate the MAR.ia reconstruction of extreme heatwaves, storms and heavy rainfall associated with several daily historical events in Belgium and compare these results with those obtained on average over the testing period.
This evaluation enables us to critically assess the ability of deep generative models, and more precisely diffusion models approaches, to faithfully reconstruct out-of-distribution events.
Doutreloup, S., Wyard, C., Amory, C., Kittel, C., Erpicum, M., and Fettweis, X. (2019). Sensitivity to Convective Schemes on Precipitation Simulated by the Regional Climate Model MAR over Belgium (1987–2017), Atmosphere, 10, 34. https://doi.org/10.3390/atmos10010034.
How to cite: Peters, S., Faulx, E., Fettweis, X., and Louppe, G.: MAR.ia: How diffusion-based approaches can reproduce extreme weather events, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18232, https://doi.org/10.5194/egusphere-egu26-18232, 2026.
Extreme value theory provides effective approaches and methods for estimating return levels RL (with a typical return period >100 years) of extreme events. However, the lack of sufficiently representative observations to properly fit extreme value distributions (EVDs) is a recurring problem for any metocean engineer in situations where the number of observations is limited or of poor quality [1]. To overcome this problem, augmenting the set of observations with complementary information sources is an interesting option. In this paper, we address this problem by fitting EVDs to both observations and predictions from machine learning models using the approach developed by [2]. By design, however, the predictions of machine learning models are uncertain because they are learned from a limited number of training samples. We therefore propose to explicitly take this error into account when inferring the EVD parameters within an approximate Bayesian computation (ABC) scheme combined with the Wasserstein distance [3].
The added value of this ML approach, which takes prediction uncertainty into account, is shown for cyclone-induced waves in Guadeloupe (Lesser Antilles) using a large database of extreme waves (representative of 1,000 years of cyclonic activity) that were numerically calculated within [4]. A random forest (RF) regression model is trained to link cyclone characteristics (radius, atmospheric pressure, distance to the eye of the hurricane) to significant wave height, and the quantile variant of the RF model is then used to model prediction error within the ABC scheme. Comparison with the 100-year and 500-year RL reference solutions (calculated using the complete database) shows that the ML-based approach results in low bias and high reliability of RL estimates as well as gain in computational efficiency, even when the sample size is reduced by a factor up to 10 and even when the RF prediction error remains moderate with cross-validation coefficient of determination of 70–75%. The benefit of integrating the ML prediction error is shown in different contexts, both along Guadeloupe coasts and in deep ocean environments.
[1] Jonathan et al. (2021). Ocean Engineering, doi:10.1016/j.oceaneng.2020.107725
[2] Rohmer et al. (2023). Ocean Modelling, doi:10.1016/j.ocemod.2023.102275
[3] Bernton et al. (2019). Journal of the Royal Statistical Society Series B, https://doi.org/10.1111/rssb.12312
[4] Interreg Carib-Coast program, https://www.carib-coast.com/en/
How to cite: Rohmer, J., Filippini, A. G., and Pedreros, R.: Improved return level estimates of cyclone-induced extreme waves by combining extreme value distribution and probabilistic machine learning predictions, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4874, https://doi.org/10.5194/egusphere-egu26-4874, 2026.
Tropical cyclone (TC) is one of the most hazardous and extreme weather events that permanently affect lives of all forms with increased severity over densely populated coastal regions. For decades, numerical weather prediction (NWP) models that solve complex mathematical equations to predict TC properties such as genesis, intensity and track, have been used with good effect. Due to climate change, TCs are set to become more frequent and intense, greatly endangering human lives and affecting biodiversity along the coastal regions. Thereby, multi-modal forecasts along with NWP predictions and strategic dissemination of information amongst the masses is required. Deep Learning (DL) models are yielding very good results across multiple domains on unstructured data including time series. Consequently, DL techniques are being developed to forecast various aspects of TCs too. In the current work, INSAT-3D satellite imagery in thermal infrared band TIR1 from Meteorological and Oceanographic Satellite Data Archival Centre (MOSDAC), Government of India, and best track data from India Meteorological Department (IMD) of 64 TCs that occurred over Bay of Bengal (BoB) from 2013 to 2023 are used to model the intensity and track. Intensity of TCs is represented using estimated central pressure (ECP) and maximum sustained surface wind speed (MSW) and tracks of TCs are represented using latitude (LAT) and longitude (LON) of the centre of the TCs. These data are collected from IMD annual reports. Since the INSAT-3D data represent satellite image time series, traditional Convolution Neural Network (CNN) alone would not suffice. A two-branch DL architecture based on Long Short-Term Memory (LSTM) (for processing intensity and track) and Convolution LSTM (ConvLSTM) (for processing the time series of satellite images) algorithms is modelled on the available data to obtain simultaneous short-term forecasting of both intensity and track of TCs. The best model predicts intensity with an error of 4.68±1.95 knots and 3.45±0.38 hPa and track with an error of 169.58±48.02 km for a lead time of six hours. However, the INSAT-3D data contains missing images for a large number of timestamps. A sub-field of DL known as generative artificial intelligence (GenAI) has excelled in generating new data from existing data. The fractured MOSDAC dataset is repaired to a large extent using a hybrid ConvLSTM-CNN architecture by generating images at the timestamps where satellite observations are unavailable. All gaps of 1-3 images are filled using this technique. The images are generated with an average structural similarity index measure (SSIM) of 0.96 and an average peak signal to noise ratio of 30.42 dB. The new augmented dataset is modelled for forecasting the intensity and track of TCs using the earlier architecture. The results improved significantly to give intensity with an error of 2.86±2.00 knots and 3.03±1.84 hPa and track with an error of 31.01±11.35 km for a lead time of six hours. Additionally, experiments for longer lead times also could be conducted. Thus, given a high-quality dataset, TC intensity and track can be forecast with good levels of accuracy and can be used to supplement the forecasts of traditional numerical techniques.
How to cite: Das, U., Pal, S., and Bandyopadhyay, O.: GenAI-assisted Intensity and Track Forecasting of Tropical Cyclones in Bay of Bengal using a hybrid Deep Learning architecture, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19628, https://doi.org/10.5194/egusphere-egu26-19628, 2026.
Masked Token Models (MTMs) are a highly efficient paradigm for pre-training large-scale models in video and language domains. Designed to learn representations on inherently sparse or strongly subsampled data, MTMs can be a promising choice for weather and climate prediction over long horizons. Despite their advantageous design properties these models have not yet found widespread adoption in climate science. We partly attribute this to limitations of the prevalent choice to use masking strategies that are uniform over time, which biases the learned representations toward spatial interpolation rather than predictive dynamics.
By defining a time-aware prior over the masking distribution, we are able to control this bias in a principled manner, thereby elevating the forecasting capability of MTMs to be on par with other approaches while retaining their efficiency and flexibility in adapting to multiple downstream tasks. Furthermore, we show that the choice of prior has a strong effect on the predicted uncertainty, leading to substantial improvements in terms of calibration.
As an illustrative example we train MTMs to predict the El Niño–Southern Oscillation (ENSO)—a primary driver of inter-seasonal climate variability with extreme weather impacts across the globe. Our approach yields state-of-the-art probabilistic forecasts of the tropical Pacific up to 24 months ahead and produces uncertainty estimates with an almost perfect spread-to-skill ratio over the full horizon. The strong performance on both climate model simulations and observational datasets demonstrates that MTMs can be highly effective for seasonal-to-annual climate prediction.
How to cite: Thümmel, J., Ebmeier, F., Schlör, J., Ludwig, N., and Goswami, B.: Masked Token Models as a paradigm for probabilistic forecasts in weather and climate, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20210, https://doi.org/10.5194/egusphere-egu26-20210, 2026.
Rapid advances in artificial intelligence weather prediction (AIWP) have enabled AI models to potentially outperform traditional numerical weather prediction (NWP) models while requiring only a fraction of the computational resources. However, many AI forecast evaluation studies have compared models using global metrics over limited years without focusing on sector and region-specific applications. Operationally driven benchmarking is necessary to effectively deploy these models, informing both model selection and improvements for different decision-making needs. Such benchmarking has been instrumental in driving AI progress in areas like ImageNet and AlphaFold. In this work, we benchmark the performance of six state-of-the-art AIWP models (AIFS, FuXi, FuXi-S2S, GraphCast, GenCast, NeuralGCM) and an NWP model (IFS) in forecasting local-scale agriculturally relevant monsoon onset over India. The models’ onset forecasts are compared with over a century of rain gauge–based ground truth observations, using standard verification metrics for both deterministic and probabilistic forecasts. This multiperiod evaluation is specifically designed to align with how such forecasts will be disseminated to stakeholders. In this operationally oriented benchmarking, we find that most AIWP models outperform the climatological baseline forecasts at medium-range timescales (~15 days), but exhibit comparable skill at subseasonal timescales (~30 days) in the core monsoon zone. These models also achieve comparable performance to IFS, while enabling calibration of probabilistic forecasts through precisely controlled ensembles that can be efficiently generated for multiple past decades. The speed and open-source nature of AIWPs provide the additional advantage that one can localize such models.
This benchmark guided model selection for large-scale AI-based generation and dissemination of the 2025 monsoon onset forecast to 38 million farmers in India. Our work presents a framework for developing operational, decision-oriented benchmarks that can accelerate the translation of the AI-driven second weather revolution into the democratization of weather forecasting worldwide.
How to cite: Masiwal, R., Aitken, C., Marchakitus, A., Gupta, M., Kowal, K., Pahlavan, H., Yang, T., Sun, Y. Q., Jina, A., Boos, W., and Hassanzadeh, P.: Decision-oriented benchmarking of AI weather models for subseasonal monsoon onset forecasts in India , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16083, https://doi.org/10.5194/egusphere-egu26-16083, 2026.
Subseasonal forecasts, targeting lead times from about 2 weeks to 2 months, remain challenging. This time range between medium-range weather and seasonal climate predictions is often described as a “predictability desert”, where both numerical weather prediction (NWP) models and machine learning (ML)-based systems tend to lose skill or have not been rigorously evaluated. In this work, we fine-tune a 2D Transformer-based model derived from the Pangu-Weather architecture for 30-day subseasonal forecasts. The focus is on improving week-3 and week-4 lead times by assigning extra weights to the tropics, which host slowly varying modes of variability that influence global weather. During model training, we apply region-based weighting using a smooth Gaussian function centered at the equator. This function assigns higher weights to tropical latitudes, with the width of the weighting controlled by a tunable standard deviation parameter. The model is trained on a multi-year subset of 6-hourly ERA5 reanalysis data and uses five upper-air variables (geopotential, temperature, zonal and meridional wind components, specific humidity) at 13 pressure levels, along with four surface variables (mean sea-level pressure, 2-meter temperature, 10-meter winds), totaling 69 input channels. For inference, we generate both deterministic and ensemble forecasts. The deterministic forecasts are initialized using ERA5 reanalysis fields, while the ensemble forecasts use 10 perturbed members from ECMWF’s Ensemble Data Assimilation (EDA), enabling probabilistic forecast evaluation. Forecast evaluation is conducted using both deterministic and probabilistic metrics. Compared to the 2D Transformer baseline, the fine-tuned model shows approximately 70% bias reduction and up to 50% RMSE improvement for temperature (T850 and T2m), particularly at week-3 and week-4 lead times. CRPS scores also generally improve, indicating better ensemble skill and reliability.
How to cite: Rami, S., Kieckhefen, D., Heyen, L., Debus, C., and Quinting, J.: Improving Subseasonal Weather Forecast Using Tropical Weighting: A Fine-Tuned 2D Transformer, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9693, https://doi.org/10.5194/egusphere-egu26-9693, 2026.
Posters on site: Wed, 6 May, 08:30–10:15 | Hall X5
The intensification of the frequency and severity of precipitation events has had a significant impact on densely populated urban areas, highlighting the need to improve traditional weather forecasting models. Due to the dynamic nature and interaction of atmospheric, oceanic, and terrestrial factors associated with these phenomena, forecasting these events is complex and challenging. Methods based on recurrent neural networks have surpassed traditional techniques in forecasting intense precipitation. However, challenges remain, such as measurement uncertainty and the high variability of events characterized by non-stationary phenomena. In this study, we propose a predictive model that employs recurrent neural networks trained exclusively with severe rainfall events.
The methodology developed incorporates Kriging for modeling the spatial structure of precipitation, allowing values to be estimated in locations without measurements and generating continuous rainfall fields that feed the forecast model. To capture the temporal evolution and abrupt variability associated with severe events, we use recurrent neural networks structured with sliding time windows of different sizes. This combination seeks to exploit the spatial correlation of the data and the learning capacity of time series to refine anomaly detection. The proposed approach was applied to the Metropolitan Region of Rio de Janeiro, a scenario marked by strong geomorphological complexity and high recurrence of extreme events. The results show that the integration between geostatistical interpolation and neural networks substantially improves the system's ability to capture rapid spatiotemporal variations in precipitation, which can assist risk warning systems and mitigate the socioeconomic impacts associated with these events.
How to cite: Rodrigues, D., Caseri, A., and Pesco, S.: Recurrent Neural Networks and Geostatistics Applied to the Prediction of Severe Rainfall Events and Anomaly Detection, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-647, https://doi.org/10.5194/egusphere-egu26-647, 2026.
Soil Moisture drought (SMD), characterized by insufficient soil moisture, affects water resources, crop yields, and economic stability across various temporal scales. India is an agrarian nation with ~70% of population dependent on agriculture. Forecasting SMD at sub-seasonal to seasonal (S2S) scales will support crop and water management, optimizing yields and averting losses. Traditionally, dynamical models like North American Multi Model Ensemble (NMME), CFSv2, and ECMWF's SEAS5 provide S2S predictions up to ten months, predicting drought onset and intensity. These models require post-processing through bias correction and downscaling due to uncertainties in initial conditions and parameterizations. Although dynamical forecasts show considerable skill in predicting extremes, forecast accuracy needs refinement to improve reliability and utilization in operational systems. In recent years, advancements in deep learning have shown potential to meet or surpass the quality of dynamical forecasts.
Recognizing the skill of dynamical S2S forecasts, this study develops a hybrid deep learning framework to predict SMDs in India at 1-3-month lead times. We combined dynamical forecasts from CFSv2 and SEAS5 with antecedent land-atmosphere conditions, climate drivers, and static features to predict SMDs using Graph Neural Networks (GNNs) with an extreme-aware custom loss function. GNNs have a better ability to learn spatial and temporal patterns and offer advantages over conventional models like ConvLSTM. Land-atmosphere variables include precipitation, maximum-minimum temperature, vapour pressure deficit, evapotranspiration, vegetation index, soil moisture, and wind speed. Large-scale climate drivers that influence rainfall patterns over India include El Niño, NAO, IOD, PDO, and MJO. Static features comprise soil type, position vector, elevation, and land use for essential contextual information. The model training was performed from June 1981–May 2015, and testing from June 2015–May 2022. The model performance is evaluated using metrics like probability of detection, percentage correct, false alarm rate, and equitable threat score. We also compare the model with dynamical forecasts and other benchmark deep learning algorithms to develop functional drought early warning systems.
How to cite: Verma, S. and Lanka, K.: Advanced Hybrid Deep Learning for Sub-Seasonal to Seasonal Forecasting of Soil Moisture Drought Over India, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1040, https://doi.org/10.5194/egusphere-egu26-1040, 2026.
This research aims to predict storminess in the North Sea using machine learning methods, focusing on how the stratosphere and upper troposphere influence winter storms. Understanding what drives winter storminess is essential for improving sub-seasonal prediction skill in a region strongly affected by extratropical cyclones. Using ERA5 reanalysis data (1940–2024), we built a storminess index based on storm event frequency and examined its relationship with large-scale atmospheric fields.
We predict North Sea storminess using two approaches, one based on the ACE2 climate emulator and another on the Random Forest machine learning algorithm. For the ACE2 model, we used air temperature and zonal and meridional wind patterns at 70 hPa as predictors. For the Random Forest regression model, we used December air temperature, zonal wind at 70 hPa, and geopotential height at 200 hPa as predictors. In both cases, the predictand is North Sea storminess. The ACE2 simulations show that when we add the initial conditions of years with low January storminess, with the December 2015 (selected because December 2015 was followed by a stormy January) stratospheric anomalies (colder temperatures and stronger winds), January surface wind speeds increase generally about 0.5–3 m/s across much of the North Sea. This suggests a dynamical link between early winter stratospheric conditions and stronger surface storminess. The Random Forest model combined with PCA shows a correlation of 0.55–0.60 when the predictors are from December, and the predictand is from January (December–January). When we test other month pairs, the correlation is 0.20–0.36 for November–December and January–February, but it drops to negative values (–0.44 to –0.05) for October–November and February–March. This pattern follows the seasonal cycle of the polar vortex. The circumpolar westerly jet strengthens from autumn and peaks in winter, when predictability is highest. This higher skill is likely linked to stronger stratosphere–troposphere coupling between November and January, as polar vortex anomalies develop and begin to descend toward the surface.
Overall, this research shows that stratospheric conditions play an essential role in shaping North Sea winter storminess and that machine learning methods can improve sub-seasonal predictions in this region.
How to cite: Aziz, P., Hünicke, B., Zorita, E., and Schrum, C.: Sub-seasonal prediction of storminess in the North Sea with machine learning methods, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1388, https://doi.org/10.5194/egusphere-egu26-1388, 2026.
The atmosphere is a complex, multiscale deterministic system in which processes across a wide range of spatial and temporal scales interact. Numerical weather prediction (NWP) models are designed to forecast the future state of the atmosphere. Processes operating at scales larger than a model’s grid spacing are explicitly represented through finite-difference approximations of the governing physical laws. In contrast, processes occurring at scales finer than the model’s numerical resolution cannot be explicitly resolved; their effects on the resolved scales are instead represented as a bulk forcing conditioned on the resolved state.
Traditionally, forcing from sub-grid scales is partitioned into several categories, such as convection, microphysics, and planetary boundary layer. The limitations of the physical parameterization schemes used for this purpose are well known. Although these schemes are physically motivated, they generally lack closed formulations, and their parameters must ultimately be tuned. Moreover, interactions among sub-grid physical processes, which are artificially separated into categories, remain largely unresolved. The development of such schemes is also labor-intensive. As a result, physical parameterizations have long been regarded as a major source of uncertainty in NWP models.
This study is motivated by the recognition that the influence of unresolved scales on the resolved flow is fundamentally a statistical problem. From this perspective, we seek a simple and efficient statistical framework to estimate sub-grid-scale forcing. We propose a novel approach that employs artificial intelligence (AI) to statistically emulate the combined effects of all sub-grid physical processes, rather than treating them separately as in traditional parameterization schemes. A key innovation of the proposed framework is the use of localized wavelet embedding to condition the statistical estimation of forcing on the relevant spatial scales influencing each model grid column. This wavelet-based representation captures both slowly evolving large-scale features and rapidly varying small-scale features.
In addition, a neural network (NN) model is trained to predict the difference between a dynamics-only model forecast and the corresponding verifying reanalysis. This trained NN can be interpreted as an AI-based all-physics model, as it effectively represents the stochastic effects of fine-scale processes unresolved by the NWP model on the resolved scales. By integrating information across scales and processes, this AI-based all-physics framework may enable even coarse-resolution global models to accurately simulate large-scale tropical waves arising from cross-scale interactions. This task remains challenging even for high-resolution global models. The proposed approach therefore offers a promising pathway toward more accurate and computationally efficient extended-range weather prediction.
How to cite: Chang, H.-L., Toth, Z., Tai, Y.-L., Weng, C.-K., Yang, S.-C., and Lin, P.-L.: A Wavelet-Embedded AI Framework for Unified Representation of Sub-Grid Physics in NWP, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4211, https://doi.org/10.5194/egusphere-egu26-4211, 2026.
Accurate forecasting of climate extremes such as droughts, heatwaves, and heat stress episodes at subseasonal-to-seasonal (S2S) timescales is of high importance for the public health, energy, water management, and agriculture sectors. However, there is a communis opinio that these scales, commonly referred to as the "predictability desert", represents a major scientific challenge for accurate forecasting. Indeed, despite recent progress, both state-of-the-art numerical and deep learning-based weather forecasting models still exhibit limited skill in forecasting extreme events beyond ten days (Bodnar et al., 2025; Bi et al., 2023; Chen et al., 2023; Lam et al., 2023; Chattopadhyay et al., 2020).
In this work, an alternative approach is considered by revisiting analogue forecasting methods (Marina et al., 2026; Pérez-Aracil et al., 2024). In the spirit of the K-nearest neighbor algorithm, these methods are built on the premise that atmospheric states with similar initial conditions tend to evolve in a similar manner (Zhao et al., 2016; Lorenz, 1969). As a result, they provide an interpretable and computationally efficient forecasting approach. However, the high dimensionality of the predictor space, combined with the choice of similarity metric, makes the identification of relevant analogues for forecasting extreme events non-trivial.
By drawing on architectural principles from state-of-the-art deep learning-based weather forecasting models, we propose a novel forecasting method that combines traditional analogue techniques with self-supervised learning. Global atmospheric, ocean, and land surface fields are first mapped into a low-dimensional latent space. Analogues are then identified in this learned space, enabling probabilistic reconstruction and forecasting of heat extremes. We evaluate our method in terms of analogue selection and forecast accuracy, with a particular emphasis on interpretability, physical consistency, and generalization to unseen heat extremes.
References:
Bodnar, C., et al. A Foundation Model for the Earth System. Nature, 2025.
Bi, K., et al. Accurate Medium-Range Global Weather Forecasting with 3D Neural Networks. Nature, 2023.
Chattopadhyay, A., et al. Analog Forecasting of Extreme-Causing Weather Patterns Using Deep Learning. Journal of Advances in Modeling Earth Systems, 2020.
Chen, L., et al. FuXi: a Cascade Machine Learning Forecasting System for 15-day Global Weather Forecast. Npj Climate and Atmospheric Science, 2023.
Lam, R., et al. Learning Skillful Medium-Range Global Weather Forecasting. Science, 2023.
Lorenz, E. Atmospheric Predictability as Revealed by Naturally Occurring Analogues. Atmospheric Sciences, 1969.
Marina, C. M., et al. Detection and Attribution of Heat Waves with the Multivariate Autoencoder Flow-Analogue Method (MvAE-AM). Atmospheric Research, 2026.
Pérez-Aracil, J., et al. Autoencoder-based Flow-Analogue Probabilistic Reconstruction of Heat Waves from Pressure Fields. Annals of the New York Academy of Sciences, 2024.
Zhao, Z., et al. Analog Forecasting with Dynamics-Adapted Kernels. Nonlinearity, 2016.
How to cite: Mortier, T., Decancq, C., Lemus-Cánovas, M., Insua-Costa, D., and G. Miralles, D.: A Self-Supervised Analogue Framework for Probabilistic Subseasonal Forecasting of Heat Extremes, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5045, https://doi.org/10.5194/egusphere-egu26-5045, 2026.
Austria’s expansion of renewable energy generation, together with projected increases in climate variability under climate change, is expected to substantially increase the vulnerability of the energy system to weather-driven disruptions. This challenge is particularly acute in the Alpine region, where complex topography and land–atmosphere interactions drive highly heterogeneous and rapidly evolving meteorological conditions. These Alpine-specific processes give rise to localized extreme events that are difficult to forecast and pose significant challenges for energy system operation, infrastructure planning, and grid stability.
The project EnergyProtect aims to identify present and future meteorological risk hotspots, defined as locations of renewable energy infrastructure with elevated exposure to weather conditions that can impair energy production or destabilize the electricity grid. We focus on hazardous meteorological phenomena relevant to wind and hydropower systems, including wind speed ramping, high wind and gust events, and high-precipitation episodes. Rapid wind speed changes can induce mechanical stress on wind turbines and other energy-related infrastructure, reduce operational efficiency, and trigger sudden power fluctuations that challenge grid balancing. Sustained high winds and gusts may lead to turbine cut-outs, structural damage, and pronounced power ramping events. In hydropower systems, extreme precipitation can increase tailwater levels, thereby reducing generation efficiency, while also elevating the risk of electrical faults and infrastructure damage in flood-prone areas.
The meteorological hazard assessment combines several advanced modelling approaches. Key components include (i) physics-informed machine learning techniques to detect and classify patterns of adverse weather, (ii) an ensemble of dynamically downscaled climate simulations at convection-permitting resolutions to capture Alpine-scale processes, and (iii) probabilistic estimates of event frequency, return periods, and future changes in intensity. This framework enables a consistent characterization of both present-day and future extreme weather hazards, while explicitly accounting for model and scenario uncertainty.
These meteorological datasets are subsequently integrated into a spatio-temporal exposure analysis of renewable energy assets to identify current and projected risk hotspots. We present preliminary results for multiple severity levels of wind speed and storm/gust ramping and high wind events with the potential to cause turbine cut-outs, efficiency losses, or grid destabilization. Using hourly meteorological datasets at spatial resolutions ranging from 1 to 30 km, we map the average annual occurrence of these risk events across Austria and quantify associated uncertainties. The results provide a robust basis for climate-resilient planning and adaptation strategies for Austria’s current and future energy system.
How to cite: Spiekermann, R., Schicker, I., Lexer, A., and Lehner, S.: AI-enhanced national-scale assessment of meteorological risk hotspots for wind and hydropower, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7714, https://doi.org/10.5194/egusphere-egu26-7714, 2026.
Accurately forecasting extreme precipitation remains a longstanding challenge in numerical weather prediction (NWP). Recently, data-driven Artificial Intelligence (AI) models have shown promise in improving global weather forecast accuracy, but their potential to enhance moso-scale precipitation forecasts has not been fully explored. This study evaluates the effectiveness of using forecasts from three AI models (Fuxi, Pangu, and Fengwu) compared with those from the traditional Global Forecast System (GFS) to initialize the Weather Research and Forecasting (WRF) model for simulating the extreme rainfall associated with landfalling Typhoon Bebinca (2024), the strongest typhoon to make landfall in Shanghai since 1949. A total of twenty WRF experiments were conducted across multiple initialization times, enabling a systematic and homogenized comparison of forecast performance. Results show that forecasts from the Fuxi and Pangu models provided more reliable and stable initial conditions, leading to improved predictions of typhoon track and extreme precipitation, particularly at longer lead times. Among the three AI models, Fengwu-driven simulations yielded the lowest track errors and demonstrated superior skill at shorter lead times (within 72 hours). Further physical diagnosis revealed that AI-driven WRF simulations produced more realistic thermodynamic structures, including stronger frontogenesis and enhanced convective organization, which contributed to improved rainfall forecasts. These findings underscore that high-quality large-scale initial fields from AI models not only improve the forecasts of synoptic-scale features such as typhoon track and intensity but also exert critical influence on the location and intensity of precipitation associated with mesoscale convective systems.
How to cite: Wang, R.: Improving forecasts of extreme rainfall induced by landfalling typhoon Bebinca (2024): Evaluating Fuxi, Pangu and Fengwu AI-driven WRF simulations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9078, https://doi.org/10.5194/egusphere-egu26-9078, 2026.
Despite AI-driven weather forecasting has made rapid progress, this progress has primarily focused on global models, which require processing planetary-scale data on low-resolution grids. To address this gap, given the need for high-resolution forecasts for specific regions in many research and applications, we propose a computationally efficient generative framework for short- to medium-term hourly regional forecasts. This framework ingests multi-resolution, multi-source geophysical inputs, combining 0.25° 3D atmospheric fields with 0.1° surface fields. To avoid simply stitching together heterogeneous grids, we design a coupled architecture to enable interaction between the evolving 3D atmospheric state and high-resolution surface and precipitation-related signals. The training process uses ERA5 data, satellite-derived products, and radar precipitation observations. We describe the end-to-end modeling pipeline and evaluation protocol and discuss uncertainty-aware regional forecasts achieved through generative methods.
How to cite: Ma, J.: Coupled Multi-Resolution Generative Modelling for High-Resolution Hourly Regional Weather Forecasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12019, https://doi.org/10.5194/egusphere-egu26-12019, 2026.
Tropical easterly waves (TEWs) directly impact people through wind, rain, and tropical cyclone formation in the Pacific and Atlantic Oceans. The structure and intensity of a TEW can be affected by a myriad of internal and external factors during a wave’s lifetime. Most existing statistical models of TEW intensification have been specifically designed to predict tropical cyclone formation from these waves. Understanding TEW behavior across a wide range of intensities, timescales, and geographic regions would provide insight into the general framework of TEW evolution. We use a novel TEW dataset, ERA5 reanalysis, and GridSAT brightness temperature data to train a neural network to predict vorticity and convective intensity in TEWs at lead times of 1 to 5 days over Africa, in the tropical North Atlantic, and in the eastern North Pacific. This network uses TEW-centered input data to generate a 50-member ensemble of predictions for each output variable at each lead time. We verify the network's predictive performance against forecasts from operational modeling. We identify input variables that contribute most significantly to the network’s output predictions and associated mean errors and ensemble uncertainty, and show how these findings vary for waves in different locations and of different initial strengths. Physically intuitive mechanisms seen in this investigation can help us better understand how TEWs evolve along an intensity / organization spectrum ranging from weak, dry waves to full-fledged tropical cyclones.
How to cite: Downs, W. B. and Majumdar, S.: Prediction of Tropical Easterly Waves Using Deep Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14208, https://doi.org/10.5194/egusphere-egu26-14208, 2026.
Subseasonal-to-Seasonal (S2S) forecasts, which are weather forecasts over a period spanning two weeks to two months, are challenging due to the position between short-term forecasts driven by initial conditions and seasonal forecasts governed by boundary conditions. Improving S2S forecasts skill to predict hydrological disasters like floods enables the establishment of disaster preparedness plans and reduces socioeconomic losses. Consequently, as the frequency of extreme precipitation events increases due to climate change, S2S forecasts are playing an increasingly vital role in early warning systems. However, S2S precipitation forecasts using traditional physics-based models are considered to have significant limitations due to errors arising from resolution, parameterization, and model uncertainty. Recently, interest has grown in whether data driven weather and climate models can bridge this forecasting gap.
Therefore, this study compares the precipitation forecasting performance of ECMWF and Korea Meteorological Administration (KMA) models with weather and climate foundation models to assess whether AI models can extend the predictability in the regions where S2S forecasts from traditional numerical weather prediction models are limited. Pre-trained foundation model and Multi-Source Weighted-Ensemble Precipitation (MSWEP) datasets are used for training a lightweight decoder to forecast precipitation from latent representations. We compare precipitation forecasts for nine years (2017-2025) with the MSWEP dataset, and analyze 2022 flood cases over Asia to evaluate the predictability of S2S for extreme weather events. We will show that a comparison of S2S precipitation forecast skill and extreme rainfall predictability between physics-based and AI models highlights the potential of S2S forecasts for early warning.
How to cite: Lee, E., Kim, S., Lee, D., Budamala, V., and Kim, H.: Weather and Climate Foundation Models Enhance Subseasonal-to-Seasonal (S2S) Precipitation Prediction Using Multi-Source Satellite Observations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18057, https://doi.org/10.5194/egusphere-egu26-18057, 2026.
Assessing Stochastic Interpolants for Downscaling of Climate Extremes
Erik Larsson, Ramón Fuentes-Franco, Mikhail Ivanov and Fredrik Lindsten
Assessing climate-related extremes at regional scales requires high-resolution information, typically obtained from dynamical regional climate models (RCMs). However, the computational expense of RCMs limits ensemble size and restricts the exploration of uncertainty. To address this challenge, we introduce a probabilistic machine-learning downscaling framework based on stochastic interpolants, trained to emulate 12 km HCLIM fields from coarse Earth System Model (ESM) output. By leveraging the stochastic interpolant framework, we construct a generative model that learns a direct mapping from coarse ESM inputs to high-resolution RCM simulations. This contrasts with standard diffusion-based approaches, where the model learns to transform Gaussian noise into RCM states. Our preliminary results indicate that the stochastic interpolant formulation provides a more effective and stable learning objective for the downscaling task.
A comprehensive evaluation across Europe for 1985–2014 shows that the emulator accurately reproduces the climatological distribution and magnitude of daily precipitation extremes. Maximum daily precipitation fields capture orographic and coastal hotspots seen in HCLIM, such as the Alps, western Norway, the Dinaric Alps, and the western Iberian Peninsula.
For precipitation exceeding the local 95th percentile, the emulator achieves a domain-mean Matthews Correlation Coefficient (MCC) of 0.35. It maintains stronger spatiotemporal synchronisation with the ESM than the RCM itself, with an MCC of 0.46 against EC-Earth3-Veg compared to 0.35 for HCLIM. This indicates that the emulator follows the large-scale dynamics imposed by the driving ESM, while reproducing the fine-scale intensity and spatial structure of extremes characteristic of the RCM.
For temperature extremes, skill is even higher, with MCC values exceeding 0.7 across most of Europe, confirming robust reproduction of warm-event timing and spatial extent. The emulator also correctly represents daily temperature–precipitation covariability, including the transition from positive correlations in winter to negative correlations in summer, and reproduces the geographical pattern of compound hot-dry events, although with regional biases consistent with the driving model.
Overall, these results show that the stochastic interpolant downscaling framework provides a computationally efficient pathway to generate large, high-resolution ensembles that retain ESM dynamics while delivering RCM-like representations of climate extremes, offering new opportunities for climate-risk assessment, attribution studies, and impact modelling.
How to cite: Larsson, E., Fuentes-Franco, R., Ivanov, M., and Lindsten, F.: Assessing Stochastic Interpolants for Downscaling of Climate Extremes, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21263, https://doi.org/10.5194/egusphere-egu26-21263, 2026.
Timely estimates of land surface temperature (LST) are critical in weather and climate prediction. Examples include assessing effects of extreme heat and drought on the biosphere and modelling transport processes in the atmospheric boundary layer. Yet, forecasting the spatiotemporal variability of LST remains challenging because the surface skin responds to forcing instantaneously and is controlled by multi-scale thermodynamic processes. Existing work on surface temperature forecasting largely follows two distinct paradigms: A) AI-driven and numerical weather prediction where large-scale skin temperature is simulated from Earth system models or their emulators, and B) geoscientific remote sensing where satellite-retrieved LST is extrapolated in time or space on small spatial scales, including site-scale experiments, often using statistical autoregression. While the goal of A) is to provide global estimates and atmospheric boundary conditions on coarse resolution with reduced complexity of subgrid processes, the goal of B) is often to obtain better forecasts over limited areas or local stations for downstream applications but these approaches rarely incorporate synoptic-scale meteorological context.
Large-scale approaches of medium-complexity to surface temperature forecasting that bridge these two ends and account for synoptic-scale surface meteorology while being sensitive to local land conditions remain underexplored. One reason for this is that modeling the tight coupling of spatial heterogeneity to multi-scale surface energy balance processes requires incorporation of multiple data sources at different spatiotemporal resolution. We cross these two paradigms and develop an observation-guided system that produces short-term forecasts of LST from reanalysed, coarse resolution surface meteorology and ancillary geostationary-resolution land surface properties. This system will cover the diurnal cycle and spatially larges scales at geostationary-resolution. We leverage the possibilities of multi-modal supervised learning and incorporate both reanalysis and observational data, explore memoryless and autoregressive approaches and outline opportunities to include high-resolution observations. Our approach is a first step towards effectively downscaling forecasts from the WeatherGenerator foundation model to high resolution surface conditions.
How to cite: Wesselkamp, M., Benson, V., Hoffmann, S., Zehner, M., Duveiller, G., Reimers, C., Carvalhais, N., and Reichstein, M.: Forecasting satellite-retrieved land surface temperature from reanalysis with multi-modal deep learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20891, https://doi.org/10.5194/egusphere-egu26-20891, 2026.
Climate change has intensified the frequency and severity of extreme meteorological events, placing growing pressure on the stable operation of wastewater treatment plants (WWTPs). Heavy rainfall and elevated temperatures can trigger abrupt changes in influent flow and pollutant loading, thereby challenging both hydraulic and operational stability of WWTPs. Although these responses are driven by meteorological forcing, their magnitude and manifestation differ across WWTPs. Such differences may be associated with non-climatic characteristics, including urbanization and plant capacity.
Accordingly, this study aims to evaluate the climate vulnerability of WWTPs by (1) characterizing relationships between meteorological conditions and influent dynamics and quantifying their sensitivity under extreme and non-extreme climate clusters, and (2) projecting future influent conditions under climate change scenarios using predictive deep learning models.
Daily operational and meteorological data collected from January 2016 to July 2025 were analyzed for four representative WWTPs located in a major metropolitan area in Republic of Korea. Meteorological variables were derived from Automated Weather System (AWS) observations and spatially aligned with service areas of each treatment plant. Meteorological conditions were classified using K-means clustering, and climate sensitivity was quantified by comparing extreme and non-extreme clusters using Cohen’s d effect size. Future influent conditions were projected by applying SSP5-8.5 climate scenario to a gated recurrent unit (GRU) trained on historical meteorological observations.
Meteorological clustering identified five distinct climate clusters, among which hot–wet (extreme event) conditions exerted the strongest impacts across all WWTPs. Under hot–wet conditions, influent volumes increased by approximately 38–86% relative to cool–dry clusters. In contrast, influent concentrations (mg/L) of organic matter and nutrients generally decreased by 20–40%, reflecting dilution effects. Conversely, suspended solids (SS) loads (kg/d) increased by an average of approximately 80% across WWTPs, indicating a strong linkage between rainfall and sediment transport.
In terms of treatment performance, nutrient removal efficiencies (total nitrogen (TN) and total phosphorus (TP)) declined markedly than those of organic matter and SS. Effect-size-based analysis revealed pronounced climate sensitivity, with very large effect sizes for influent flow (Cohen’s d ≈ 2.0–3.0) and consistently large sensitivities for SS load and nutrient removal (d > 1.0). In contrast, organic matter removal showed relatively smaller sensitivities. These response patterns were subsequently used to assess climate vulnerability across WWTPs with different levels of urbanization and plant capacities, highlighting substantial inter-plant variability in climate sensitivity.
Building on the climate sensitivity patterns derived from historical observations, scenario-based projections suggest that increasing frequencies of extreme weather were likely to further amplify influent variability and pollutant loading under future climate conditions. Taken together, the historical analysis demonstrates that the vulnerability of WWTPs to climate change is influenced not only by extreme weather patterns but also by intrinsic system characteristics. The scenario-based projections extend these insights by highlighting potential future risks under climate forcing. By integrating meteorological clustering, effect-size-based sensitivity analysis, and scenario-driven influent projections, this study provides a practical framework for identifying vulnerable facilities and informing climate adaptation strategies, including capacity planning, nutrient management under extreme influent conditions, and prioritization of infrastructure upgrades.
How to cite: Park, H. and Kim, Y. M.: Climate Vulnerability of Wastewater Treatment Plants to Extreme Weather: An Effect-Size-Based Sensitivity Analysis of Influent and Performance, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6433, https://doi.org/10.5194/egusphere-egu26-6433, 2026.
Under changing climate conditions, it has been observed that the frequency of extreme events has increased significantly worldwide. India has also experienced numerous flash flood events over the past two decades, leading to substantial socio-economic losses. India receives 70-80% of its annual rainfall during the southwest monsoon, which affects almost all parts of the country except the southeastern coast of Tamil Nadu. Therefore, it is crucial to improve early warning systems, especially for short-term precipitation forecasts. Various national and international organizations publish forecasts for different weather parameters, such as precipitation, temperature, wind speed, etc., derived from Numerical Weather Prediction Models (NWP); these datasets often show significant spatial and temporal biases at different lead times. In this study, the goal has been to identify the spatial and temporal biases in forecast data from NCMRWF, ECMWF, and NCEP for the years 2018 to 2023, using IMD gridded rainfall data and CMORPH-NOAA satellite data as ground truth for the southwest monsoon (June to September). For each grid, the spatial correlation has been evaluated across eight neighbouring grids and the central grid, while temporal cross-correlation has been assessed over 12-hour, 24-hour, and 48-hour lead and lag periods to determine the temporal accuracy of each NWP product for 24-hour lead times, using 00:00 UTC as the reference for both ground truth accumulation and forecasts.
This study introduces a spatio-temporal deep learning–based integration framework that combines three separate NWP rainfall forecasts into a single, skill-enhanced 24-hour prediction by explicitly considering directional spatial dependence and temporal lead–lag relationships, with particular relevance for extreme rainfall detection during the monsoon season. The methodology employs a spatio-temporal deep learning framework in which three NWP precipitation forecasts are encoded separately using direction-aware neighbourhood information and lag–lead temporal context, allowing the model to learn model-specific spatial and temporal error characteristics. These encoded features are dynamically combined through an attention-based integration mechanism to produce an optimized 24-hour rainfall forecast. The combined forecast is evaluated solely at a 24-hour lead time during the South-West Monsoon season using high-resolution rainfall observations. Results indicate that the proposed directional–temporal integration consistently outperforms all individual NWP forecasts, showing significant improvements (20-50% across various parts) in various standard error metrics, including RMSE and correlation coefficient values.
The study is expected to effectively reduce the local bias in short-term rainfall forecasts over India, ultimately leading to the development of more efficient weather forecasting technologies. Additionally, the future scope of the study aims to introduce a novel approach that combines both physics-based and AI-based predictions, with the goal of establishing a benchmark for improving India's weather forecast system.
How to cite: Majumder, A. and Narasimhan, B.: A Spatio-Temporal Deep Learning Framework for Integrating NWP Products to Improve Short-Range Monsoon Rainfall Forecasts over India, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12958, https://doi.org/10.5194/egusphere-egu26-12958, 2026.
Under the influence of global climate change and human activities, the frequency and intensity of extreme weather events—such as heavy precipitation and severe droughts—have increased markedly. Flood disasters triggered by intense rainfall have severely threatened lives, property, and regional socioeconomic development. To address the challenge of precise prevention and control of short‑duration rainstorm‑induced flash floods in the complex terrain of Northwest China, this study focuses on the Ningxia region, located within China’s arid‑semi‑arid transition zone. By integrating Water Internet technology, big data, and deep learning, we construct an intelligent flash flood disaster prevention and control system.
In rainfall forecasting, we have (1) developed a radar‑based precipitation retrieval model through data fusion and calibration, achieving a retrieval accuracy of R² > 0.75 and NMAE < 0.3; (2) proposed an attention‑mechanism‑driven radar echo extrapolation technique that attains over 80% accuracy for a 3‑hour lead time; and (3) built a rapid‑cycle, multi‑source data assimilation rainfall forecast model incorporating GNSS water vapor tomography.
For flood forecasting, we (1) introduced a forecasting technique that couples multi‑source rainfall predictions with a distributed hydrological model, yielding accuracy above 80%; and (2) constructed a runoff simulation model for mountainous basins by integrating radar and terrain data with adaptive pooling and attention mechanisms, achieving over 85% forecast accuracy.
In the domain of intelligent flood regulation, a real‑time operational model based on rainfall‑runoff forecasting has been developed. By combining flood forecasts with a simplified inundation model, the system enables large‑scale watershed flood analysis.
How to cite: zhen, Q., yuying, C., jiahua, W., and jieyu, J.: A Technical Framework for Whole-Process Forecasting of Rainstorm-Induced Flash Floods Coupling Artificial Intelligence and Physical Mechanisms, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20578, https://doi.org/10.5194/egusphere-egu26-20578, 2026.
