Hydrological modeling has entered a new era in recent years, largely driven by the curation of extensive datasets and availability of open-source machine learning libraries. While traditional physically based models have been key to improving our understanding of hydrological systems and establishing early warning systems, they often face challenges such as high computational costs and requiring simplifications of complex processes. Conversely, machine learning methods, despite potential pitfalls such as generating unphysical outputs and requiring large volumes of training data, are computationally quick and capable of capturing highly non-linear relationships. Hybrid hydrological modeling bridges these approaches, combining the efficiency and flexibility of machine learning with the proven capabilities and interpretability of traditional physical models.

This talk will provide an overview of the hybrid hydrological modeling research being conducted at the European Centre for Medium-range Weather Forecasts (ECMWF). Using case studies, we will show how machine learning methods could be incorporated into the pre-established physical modeling chain, addressing both scientific and operational challenges. Examples will cover the full modelling chain including the coupling of machine learning and physically based models, the use of emulators of sub-models to reduce computational overhead, and the integration of data-driven techniques to correct model biases in real time. The development of a machine learning forecasting model will also be discussed as a component of a hybrid multi-model system. Attention will be given to the practical aspects of implementation, including computational efficiency both for an operational system and for sensitivity and calibration experiments, scalability to large operational systems, and the potential to incorporate new datasets, such as remote sensing data, into hybrid frameworks. We will also discuss how artificial intelligence can be used to support auxiliary services such as simulation verification.

Finally, we will reflect on the implications of hybrid hydrological modeling for advancing hydrological science and operational forecasting. By combining the strengths of physical and machine learning models, this approach has the potential to improve flood prediction, water resource management, and climate impact assessments. This hybrid approach marks an important step forward in the development of hydrological modeling, enabling more accurate, efficient, and actionable understanding of water systems in a rapidly changing world.

How to cite: Matthews, G., Baugh, C., Chantry, M., Mazzetti, C., Mosaffa, H., Pinnington, E., Raoult, N., Ruparell, K., Taccari, M.-L., Pappenberger, F., and Prudhomme, C.: A new era of hybrid hydrological forecasting at ECMWF , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11925, https://doi.org/10.5194/egusphere-egu25-11925, 2025.

Accurate streamflow forecasting in both managed and natural catchments is critical for sustainable water resource management in the UK. Simulating hydrological processes such as flashy flood responses and reservoir-influenced flow dynamics remains a significant challenge, particularly in non-natural catchments where human interventions alter natural flow regimes. This study evaluated the effectiveness of three modelling frameworks across 341 selected UK catchments from the CAMELS-GB database: the HBV (a conceptual hydrological model), Long Short-Term Memory (LSTM, a data-driven model), and a hybrid Physics-Informed Machine Learning (PIML) model supplemented with hydrological signatures. The regional and spatial patterns of their performance was investigated using evaluation metrics such as Nash-Sutcliffe Efficiency (NSE) and Kling-Gupta Efficiency (KGE) during both the calibration and validation phases. The results show that LSTM outperforms HBV in 65% of the catchments, particularly in northern Scotland and the western UK, which have steep terrain with rapid runoff; but it lacks the physical interpretability that HBV provides. Despite its advantages in natural catchments, HBV tends to produce unsatisfactory simulations for   processes such as snowmelt and rapid storm response, as well as for regulated flows in reservoir-affected catchments, which are common in central and southern England. The incorporation of hydrological signatures (e.g., baseflow index, rainfall-runoff ratio) into the PIML framework addresses this limitation by encapsulating key reservoir processes (e.g., flow smoothing, seasonal redistribution) and anthropogenic influences, allowing for improved streamflow predictions and enhanced interpretability. Our study emphasises the need for hybrid modelling approaches that combine the physical coherence of conceptual models with the adaptability of data-driven procedures.  The findings highlight the necessity of adapting models to local conditions and accounting for the effects of human activity, providing a reliable way to enhance streamflow predictions in complicated and regulated catchments.

How to cite: Moon, R., Han, S., Graham, L., and Hannah, D.: Bridging Physics and Machine Learning: A Signature-enhanced Hybrid Framework for Streamflow Prediction in Complex Catchments, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12113, https://doi.org/10.5194/egusphere-egu25-12113, 2025.

Accurate streamflow prediction is crucial for water resources management, particularly in the regions facing challenges such as water scarcity and hydrological unpredictability. Physical-based hydrological models have long been used for rainfall-runoff simulations by solving equations governing hydrological processes in a typical watershed. In addition, Machine Learning (ML) models emerged as versatile data-driven approaches capable of capturing intricate patterns of hydroclimatic variables, which can be used for streamflow prediction.

The aim of this study is to compare performances of two distinct approaches: (i) the process-based and semi-distributed Hydrological Predictions for the Environment (HYPE) model and Extreme Gradient Boosting (XGBoost), a tree-based ML algorithm. The case study is upper Reno River Basin, situated in northern Italy. Precipitation across the basin varies considerably due to orographic influences, while this spatial variability drives diverse seasonal and regional streamflow patterns. For this purpose, a 5-km gridded meteorological data (the ERG5 dataset) was used as input for both models from 2001 to 2022. The database was developed by ARPAe-SIMC for the Emilia-Romagna region in Italy. Furthermore, the streamflow was considered as output results. For the sake of comparison, both models were calibrated using the same time series, partitioning the data into 75% for calibration/training and 25% for testing.

The simulation performance for river discharge showed high values of the Kling-Gupta Efficiency (KGE) for the training phase as XGBoost showed slightly better values of KGE (0.86) than that of HYPE (0.82). For the test period, KGE around 0.8 was obtained by both models. Thus, the KGE values were comparable for both models, with HYPE slightly outperforming XGBoost (0.82 vs. 0.78). The flow-duration curves revealed that both models performed well for estimating peak discharges (below 30% occurrence). However, for drier conditions, HYPE shows a better agreement with the observed data, while ML tended to overestimate it.

The results indicate that traditional hydrological models performed slightly better than XGBoost for streamflow estimation in the region under investigation. The performance of XGBoost may be improved if seasonality was taken into account, which can be explored in future works. Based on the comparative analysis, ML techniques can provide a suitable alternative in cases where little is known about the region’s hydrological characteristics, leveraging data patterns without requiring detailed process knowledge. Nonetheless, the application of ML requires caution, as its black-box nature may obscure the underlying physical and hydrological processes, potentially leading to misinterpretation of results. Finally, this comparison provides valuable guidance for researchers and practitioners in selecting appropriate tools for streamflow prediction tasks.

Acknowledgements: This research work was carried out as part of the TRANSCEND project with funding received from the European Union Horizon Europe Research and Innovation Programme under Grant Agreement No. 10108411.

How to cite: Niazkar, M., Cenobio-Cruz, O., Mozzi, G., Di Baldassarre, G., and Pal, J.: Hydrological Modelling vs. Machine Learning for Water Availability: Case Study from the Reno Basin (Italy), EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17296, https://doi.org/10.5194/egusphere-egu25-17296, 2025.

Despite the advances in hydro-meteorological forecasting systems, challenges to accurately forecast streamflow at seasonal time horizons still remain, especially when models are applied across a strong hydro-climatic gradient. In this work, we explore the potential of AI-based approaches combined with the output of process-based hydrological models and meteorological forecasts from Numerical Weather Prediction models to enhance seasonal streamflow forecasts, with lead-times up to 30 weeks. We employ the multi-time scale Long Short-Term Memory (MTS-LSTM) model trained with a combination of reanalysis data from the process-based pan-European E-HYPE hydrological model, in-situ observations from GRDC, and bias-adjusted seasonal meteorological forecasts from the ECMWF SEAS5 prediction system. The MTS-LSTM is developed at the pan-European scale, using more than 500 catchments over Europe, which lie in 11 different clusters according to their hydrological regime. We then compare the AI-based forecast performance against streamflow climatology and the E-HYPE streamflow forecasts. Our results show that the streamflow forecasts based on MTS-LSTM outperform the E-HYPE ones in catchments characterised by highly variable and flashy hydrological response and snow-dominated catchments with high seasonality. However, the MTS-LSTM underperform compared to E-HYPE results in catchments with highly variable streamflow regimes and long recessions. These preliminary findings highlight the potential of AI approaches to enhance streamflow predictability at seasonal lead-times across Europe’s strong hydro-climatic gradient, having both scientific and operational added value.

How to cite: Bertini, C., Du, Y., van Andel, S. J., and Pechlivanidis, I.: AI-based seasonal streamflow forecasts across Europe’s hydro-climatic gradient, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10567, https://doi.org/10.5194/egusphere-egu25-10567, 2025.

Hydrological modeling provides a quantitative foundation for effective water resource management. Simulating discharge from meteorological forecasts is essential for flood prediction and risk assessment.

Traditional hydrological models, such as the Hydrological Predictions for the Environment (HYPE) model, leverage explicit equations to represent well-known catchment characteristics and provide process-based discharge forecasts. However, these models often struggle to capture unknown or poorly understood spatio-temporal dependencies and nonlinear dynamics.

In contrast, machine learning approaches, such as Long Short-Term Memory (LSTM) networks, are able to learn complex patterns without requiring pre-defined relationships. However, these networks introduce variabilities into hydrological model simulations, which in turn complicates the development of well-supported arguments based on their findings.

A promising solution is a hybrid model in which the physical model's output serves as dynamic input to the LSTM. This approach preserves the strengths of physics-based models in representing well-understood hydrological processes while allowing the LSTM to capture implicit dependencies.

This study investigated the application of hybrid hydrological modeling for simulating discharge in Danish catchments by combining simulated discharge from the Danish HYPE model (DK-HYPE) with an LSTM model. The analysis encompassed 570 catchments, characterized by static attributes and dynamic variables. Dynamic variables were derived from a high-resolution CAMELS dataset (DK-CAMELS) with a spatial resolution of 1x1 km, from Danish weather stations covering the time period from 2001 to 2022.

Fifteen LSTM models were trained under various configurations: different sequence lengths (30, 90, and 365 days), inclusion of static attributes, and utilization of DK-HYPE outputs. Model training used two loss functions—Mean Squared Error (MSE) and the Nash-Sutcliffe model efficiency coefficient (NSE)—while performance was evaluated using the Kling-Gupta Efficiency (KGE), Flow Balance (FBAL), and Critical Success Index (CSI).

Incorporating static attributes enhanced model accuracy, while longer sequence lengths captured hydrological dependencies. Across all configurations, the LSTM models outperformed DK-HYPE. The best-performing hybrid model achieved a KGE of 0.7 and a CSI of 0.36, a significant improvement over DK-HYPE's baseline values of 0.01 for KGE and 0.18 for CSI. Similarly, the standalone LSTM model, which excluded DK-HYPE outputs during training, achieved a KGE of 0.71 and a CSI of 0.35.

While the hybrid model did not demonstrate a clear advantage over the pure LSTM model with longer sequence lengths, it outperformed the pure LSTM model with shorter sequence lengths. Specifically, comparing models using sequence length of 30 days, the hybrid model achieved a KGE of 0.65 and a CSI of 0.36, compared to the pure LSTM model's KGE of 0.59 and CSI of 0.31, which is most likely because it utilized information from DK-HYPE.

This project is a step towards combining physics-based models with data-driven models for the national flood warning system. Further work should focus on fine-tuning the hybrid models and integrate them into an ensemble towards building robust systems for flood forecasting.

How to cite: Jensen, L. D., Martinsen, G., Madsen, H., Aarestrup, P., and Payet-Burin, R.: Utilizing Physics-Based Predictions as Inputs to LSTM Models for Robust Data-Driven Discharge Simulations of Gauged Catchments Across Denmark, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10870, https://doi.org/10.5194/egusphere-egu25-10870, 2025.

Accurate streamflow predictions in glacial basins are critical for effective water resource management and flood risk mitigation, especially under changing climatic conditions. This study presents a hybrid modeling framework that integrates the SWAT-GL model, an extension of the Soil and Water Assessment Tool (SWAT) designed to include glacial hydrological processes, with advanced machine learning techniques (Random Forest, Support Vector Regression, and Multilayer Perceptron).

SWAT-GL was selected for its proven ability to simulate glacial hydrological processes at a daily scale. However, its application at an hourly scale is limited due to the reliance on the Green-Ampt infiltration method, which is less suitable for representing the unique soil characteristics typically observed in glacier-fed basins. To overcome this limitation, machine learning models were employed to refine the daily SWAT-GL outputs into hourly predictions. An ensemble model was developed to enhance accuracy, combining the complementary strengths of the individual machine learning approaches.

The model was calibrated using data from 2021 to 2023, with a one-year warmup period (2020), and validated with observed data from January 2024 to September 2024. Meteorological forecasts from ECMWF-IFS and MOLOCH models were incorporated, providing hourly data on precipitation, temperature, solar radiation, and wind speed. This approach enabled day-ahead operational forecasting, aligning model outputs with real-time management needs.

The ensemble model showed strong performance during training and testing, highlighting its robustness in refining daily SWAT-GL outputs into accurate hourly predictions.

The hybrid framework was applied to the Forni Basin, a glacier-fed system in the Italian Alps characterized by high variability in meltwater contributions and limited hydrological data. By addressing key challenges such as input uncertainties, limitations of process-based modeling at sub-daily scales, and scaling from daily to hourly forecasts, the model offers a robust tool for predicting streamflow in data-scarce, glacierized regions. This study highlights the potential of hybrid approaches to improve hydrological forecasting accuracy and scalability, contributing to the sustainable management of water resources in sensitive alpine environments.

How to cite: Malago', A., Schaffhauser, T., Bouraoui, F., Lazzarini, P., Marziali, A., Ravasi, A., and Bertelli, V.: Hybrid Hydrological Modeling in the Forni Basin: Combining SWAT-GL and Machine Learning Techniques, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3263, https://doi.org/10.5194/egusphere-egu25-3263, 2025.

Recent advances in Earth observation and data collection technologies have made high-resolution hydrological datasets increasingly accessible, enhancing our capabilities for monitoring and predicting hydrological processes. While a variety of artificial intelligence (AI) models can be employed to leverage these datasets, the challenges and opportunities of different AI approaches in the context of high-resolution data availability remain only partially explored. For instance, although Long Short-Term Memory (LSTM) networks are widely used for discharge prediction, the potential of Graph Neural Networks (GNNs)—which naturally represent river networks as graphs and capture spatial dependencies—has yet to be fully investigated.

In this study, we conduct a comprehensive analysis of GNN-based models for river discharge prediction in the Danube River Basin. Leveraging the LamaH-CE (Large-Sample Data for Hydrology and Environmental Sciences for Central Europe) dataset, we incorporate both dynamic features (e.g., daily precipitation, temperature, soil moisture) and static variables (e.g., digital elevation model, river density, basin area). Three architectures—GNN, a hybrid LSTM-GNN, and a standalone LSTM—are trained, validated, and tested at daily time steps from 2000 to 2017.

We further investigate the impact of network density and high-resolution (1km) soil moisture and precipitation data on discharge prediction accuracy. Our analysis reveals the potential advantages and limitations of these architectural approaches in river discharge prediction under high-resolution data availability and underscores the growing importance of harnessing graph-based deep learning methods for hydrological applications.

How to cite: Mosaffa, H., Prudhomme, C., Chantry, M., Stephens, L., Rüdiger, C., Pappenberger, F., and Cloke, H.: A Comprehensive Analysis of Graph Neural Networks for River Discharge Prediction: High-Resolution Applications in the Danube Basin, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13016, https://doi.org/10.5194/egusphere-egu25-13016, 2025.

The Jhelum River, which is the major river of the signal, contributes to the Indus River system as one of the notable tributaries and is bestowed with crucial importance in adhering its water for various uses, including irrigation, hydropower supply and domestic purposes. However, it is very vulnerable to serious floods that cause many losses of life and property. As a result, precise flood forecasting in the Jhelum River is essential to facilitate proper disaster response and prevention strategies. Flood forecasting is critical to disaster preparedness, particularly in countries vulnerable to repeated hydrologic disasters. This research aims to improve the flood forecasting technique applicable to the Kupwara district of Jammu and Kashmir, as the area is frequently ravaged by floods, mostly occasioned by its geographical and climatic attributes. We employ hydrometeorological data from January 1975 to December 2023 to investigate the interaction of the factors that determine flood occurrence, and we evaluate the capability of data-based models in providing reliable monthly flood predictions. This study proposes Support Vector Machines (SVMs), Artificial Neural Networks (ANNs), and finally, the hybrid model SVM-PSO (Particle Swarm Optimization) to predict flood events in the Jhelum River. The results of the hybrid SVM-PSO model show maximum goodness of fit with an R² value of 0.9562, minimum MSE value of 9.2237 and Nash-Sutcliffe Model Efficiency of 0.9483. These outcomes illustrate the model's strengths for flood forecasting for Kupwara; its application to disaster risk reduction is valuable. The study’s findings highlight the possibility of extending the applications of progressive AI tools to reduce the effects of flooding and preserve the areas’ susceptible populations and assets in the Kupwara district.

This research was supported by the Empowerment and Equity Opportunities for Excellence (EEQ) in Science (Dr SS) under SERB, Govt. of India, under grant no. EEQ/2023/000585

How to cite: Hamid, H. and Samantaray, S.: Hybrid SVM-PSO Approach for Flood Prediction in the Jhelum River Basin: A Case Study of Kupwara District, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-608, https://doi.org/10.5194/egusphere-egu25-608, 2025.

We propose a robust framework for flood forecasting that integrates advanced hydrological modeling with high-performance computing to deliver accurate and timely predictions. At its core, the India Land Data Assimilation System (ILDAS) leverages multiple Land Surface Models (LSMs) and routing models to enhance flood prediction capabilities. The framework utilizes the Noah-MP LSM for simulating land surface processes, with runoff routed using mizuRoute, a vector-based hydrodynamic model, to estimate critical flood metrics. Calibration is optimized through an HPC-enabled scheme, ensuring precise model tuning. The system ingests meteorological forecasts from the National Center for Medium-Range Weather Forecasting (NCMRWF), ISRO, and GEFS, offering lead times of 1 to 10 days. Operating at a spatial resolution of 12 meters, it delivers nationwide flood forecasts within 15 minutes, with outputs available at daily and sub-daily temporal resolutions. The flood forecast is improved with an ML based hydrological post-processor to enhance the forecast information. A case study focusing on the flood-prone Brahmaputra basin demonstrates the system's effectiveness in accurately predicting flood events, underscoring its potential as a valuable operational tool for flood preparedness and risk mitigation. 

How to cite: Deka, P. and Saharia, M.: A Medium-Range Ensemble Flood Forecasting System for India., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-921, https://doi.org/10.5194/egusphere-egu25-921, 2025.

Flood early warning systems rely on accurate streamflow forecasts. Deep learning based approaches have been widely shown to outperform traditional, process-based approaches. While literature is rich with comparisons between these opposing modelling paradigms, most comparisons have been conducted at daily temporal resolutions and feature spatially coarse (i.e., lumped) process-based models. Flood forecasting applications, especially those in flashy urban catchments, rely on sub-daily forecasts. In this work, we compare the performance of a state-of-the-art regionally trained LSTM models with semi-distributed StormWater Management Models (SWMM) at temporal frequencies ranging from 15-minutes to 1-day, for roughly 40 highly urbanised catchments in Toronto, Canada. Results show that the LSTM approaches struggle at fine temporal resolution and when limited observed data is available. In contrast, SWMM models can be automatically parameterized and calibrated using comparatively much less data. While the amount of available historical data would be enough to train deep learning models at a daily resolution, it is insufficient to train hourly models, which we attribute to the comparatively more complex urban rainfall-runoff system. Potential solutions to this problem include model transfer between space and different temporal frequencies. Finally, another contribution of this work is LSTM hyperparameter optimization, which is not widely documented at a sub-hourly resolution. Results from this research reaffirm the need for multi-model approaches for flood forecasting, particularly in urbanised catchments.

How to cite: Snieder, E. and Khan, U.: Comparing deep-learning and semi-distributed models for flow forecasting at fine spatial and temporal resolutions: a case study of 40 urbanized catchments in Toronto, Canada., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13060, https://doi.org/10.5194/egusphere-egu25-13060, 2025.

Use of complex, high-resolution integrated hydrologic models offer the most comprehensive and detailed representations of groundwater, surface water, and land surface processes, but are challenging to use for forecasting tasks due to high computational costs and parameter uncertainty. On the flipside, machine learning approaches are highly accurate and can be computationally frugal for targeted tasks, but are difficult to audit and must be retrained to adapt to new tasks or domains.

In this work we present several case studies of using deep learning surrogate modeling approaches for integrated hydrologic modeling that alleviates many of the weaknesses of taking a purely physically based or purely data driven approach. We first show how deep learning surrogates can readily achieve orders of magnitude speedup over the original physically based models with high degree of accuracy, which allows for on demand forecasting. While this approach is great for generating forecasts from the original model configuration, it is still challenging to adapt to new scenarios such as use in parameter calibration or running long simulations such as climate change scenarios. We close the presentation by discussing recent work to address these challenges using model inversion techniques and by developing hybrid model emulation strategies.

How to cite: Bennett, A., Triplett, A., Melchior, P., Maxwell, R., and Condon, L.: Surrogate modeling for large-scale integrated hydrologic modeling: A case study in deep learning, model inversion, and hybrid methods across the Continental United States, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13539, https://doi.org/10.5194/egusphere-egu25-13539, 2025.

Groundwater depletion in the Bist-Doab region of the Punjab State of India is a significant threat to sustainable agricultural practices, underscoring the need for effective management strategies. Modelling groundwater heads is essential for understanding groundwater flow dynamics, trends, and their interaction with surface water. It helps assess the aquifer's health, prevent over-extraction and contamination, and predict ambient groundwater responses to extreme events such as droughts or floods. Inaccurate groundwater models, which overestimate or underestimate groundwater levels and fail to capture temporal fluctuations, hinder proper water management. These errors lead to suboptimal decisions regarding water allocation and resource sustainability and ultimately impact crop yields and water availability.

This study aims to integrate physically-based models, such as those developed by MODFLOW, with machine-learning algorithms to improve prediction accuracy and support more informed decision-making. MODFLOW was used to simulate groundwater flow under both steady-state and transient conditions, utilizing field hydrogeological data from existing literature. Machine learning (ML) models, including Support Vector Regression (SVR), Random Forest (RF), and Artificial Neural Networks (ANN), were trained and tested on historical groundwater levels and meteorological data to enhance prediction accuracy.

The methodology employs data-driven models (DDMs) as error-correcting tools for the physically-based models. Historical residuals, calculated as the difference between observed and simulated groundwater heads, were used as inputs alongside features such as well location coordinates, simulated groundwater heads, and time of measurements. ML techniques such as SVR, RF and ANN were used to train the DDMs, which learn systematic errors in the physically-based model by analysing these historical residuals. Outputs include predicted systematic errors and updated groundwater heads, where corrections are applied to the initial simulated values. The effectiveness of the DDMs relies on the structure and patterns of the residuals in the physically-based model, with strong correlations between the groundwater heads, leading to better error correction and improved predictive accuracy.

Results show that integrating MODFLOW with ML, significantly reduces model error compared to traditional simulation approaches. The combined model effectively captures both seasonal fluctuations and long-term trends in groundwater levels, leading to more accurate predictions. The developed framework provides a reliable tool for improving groundwater resource management and optimizing water allocation strategies, ultimately supporting the sustainable management of groundwater in agriculturally stressed regions like Bist-Doab.

How to cite: Kantode, A., Prashanth, T., and Ganguly, S.: Development of a precise regional-scale groundwater model by coupling MODFLOW & Machine Learning algorithms: A case study in Bist-Doab region, Punjab, India, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1365, https://doi.org/10.5194/egusphere-egu25-1365, 2025.

Reliable forecasting of resource availability and demand evolution is paramount for effective community planning. The forecast horizon, which can range from a few days to several years, plays a crucial role in this process. Short-term forecasts allow for the optimization of withdrawals according to demand, while medium-term forecasts enable anticipation of resource scarcity risks, planning operations on sensitive structures, or anticipating demand peaks. Long-term forecasts, on the other hand, help anticipate and analyze development strategies and usage scenarios.
 
This study, a component of the Water Resources Forecast (WRF) SUEZ’s project, partially funded by the French Ecological Transition Agency (ADEME’s innov’eau initiative), introduces an innovative approach to predicting river flow rates. We assess the performance of three distinct model types: traditional lumped rainfall–runoff conceptual models with two reservoirs, classic AI models (Random Forest, LSTM, etc.), and hybrid models that synergize AI and conceptual models for enhanced predictive accuracy.
 
Preliminary findings suggest LSTM and that hybrid models utilizing LSTM demonstrate superior short-term performance, reducing error rates by 41% compared to standalone conceptual models. These results indicate the potential of AI and hybrid models in improving the accuracy of resource availability forecasts. The analysis of the medium and long-term performances of the forecast models is currently underway and the findings will be presented at the conference.
 
This ongoing research contributes to the development of Aquadvanced® Water Resources, a comprehensive platform aimed at monitoring and forecasting various resource types, both underground and surface. By aligning resource availability with water demand forecasts, this tool will facilitate resource management strategies.

How to cite: Shakil, A., Sakarovitch, C., Bouhafa, N., and Leclerc, C.: Advancing Resource Forecasting: An Evaluation of Hybrid Predictive Models on River Flow Rates, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16839, https://doi.org/10.5194/egusphere-egu25-16839, 2025.

The Segura River Basin, which supplies water for agriculture, receives water from the Upper Tagus River Basin (UTRB) through the Tagus-Segura water transfer, involving two reservoirs: Entrepeñas and Buendía. Accurate reservoir inflow forecasts, particularly seasonal ones, are crucial for making better and more reliable water transfer decisions. This study introduces a methodology for seasonal forecasting using ensemble weather forecasts from climate models, with a focus on the SEAS5 model from the European Centre for Medium-Range Weather Forecasts (ECMWF). Initially, by combining the global climate model with machine learning algorithms, bias correction of daily precipitation and temperature forecasts is achieved. The Soil and Water Assessment Tool (SWAT+) hydrological model is employed to simulate inflows to the Entrepeñas and Buendía reservoirs, calibrated against observed inflows. The first five years from 1995 to 1999 are used for warm-up, the period from 2000 to 2009 for calibration, and from 2010 to 2019 for validation. The calibrated SWAT+ model is then forced with bias-corrected meteorological data forecasts to predict reservoir inflows for the upcoming months. The SWAT+ model's performance during calibration and validation was very good, with monthly NSE values exceeding 0.7 and PBIAS values below 14% for both reservoirs. When the model was forced with bias-corrected hydrological forecasts, it performed well, demonstrating the effectiveness of bias-corrected forecasted meteorological data in predicting reservoir inflows. This work was supported by the Spanish Ministry of Science and Innovation, under grants PID2021-128126OA-I00.

Keywords: SWAT+, machine learning, coupled modelling, streamflow simulation, seasonal hydrological forecasting

How to cite: Senent-Aparicio, J., Jimeno-Sáez, P., Asadi, S., Bilbao-Barrenetxea, N., Castellanos-Osorio, G., López-Ballesteros, A., and Cabezas, F.: Coupling SWAT+ and machine learning-Enhanced global climate models for seasonal hydrological prediction, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19297, https://doi.org/10.5194/egusphere-egu25-19297, 2025.