Floods and droughts in the Upper Blue Nile Basin pose persistent challenges to water security, agriculture, and transboundary water management across Ethiopia, Sudan, and Egypt. Historical records indicate that extreme flood and drought events recur frequently in the basin, with major floods recorded in several years over the past decades (e.g., 1988, 1998, and 2014), while severe drought conditions have affected a substantial proportion of years, notably during periods such as 1984, 2002, and 2015. These recurring hydrological extremes underscore the severity of climate variability in the basin and highlight the urgent need for reliable forecasting frameworks. This study develops a hybrid hydrological–artificial intelligence (AI) framework in which a physically based hydrological model (HEC-HMS) is first used to simulate river discharge, and its outputs are subsequently post-processed using data-driven AI forecasting techniques to improve discharge prediction under data-limited conditions. The methodology combines GIS-derived catchment characteristics, observed discharge records (2000–2006), and three AI approaches Linear Regression, Exponential Smoothing (ETS), and Random Forest Regression applied at the El-Deim gauging station. HEC-HMS was first calibrated across thirteen sub-catchments between Lake Tana and El-Deim to establish a physically grounded baseline simulation. AI models were subsequently trained and validated to capture nonlinear behavior and seasonal variability in river discharge. The 2000–2006 calibration period was chosen due to limited availability of continuous discharge data. The selected AI models represent increasing complexity: Linear Regression as a baseline, ETS for capturing seasonality in short time series, and Random Forest for nonlinear relationships, while avoiding the high data demands of deep learning models such as LSTM or ANN. Results demonstrate that the hybrid framework significantly imshows forecasting performance. A paired t-test on RMSE values confirmed that the observed error reduction is statistically significant. Among the tested models, ETS achieved the highest accuracy, reducing the root mean square error (RMSE) from approximately from approximately 4 m³ s− 1 for HEC-HMS to about 1.4 m³ s− 1. Calibration performance was evaluated using standard statistical metrics, including the Nash–Sutcliffe Efficiency (NSE) and the coefficient of determination (R²), which indicated good agreement between simulated and observed discharges across the thirteen sub-catchments. Long-term projections (2018–2030) indicate increasing wet-season flows with pronounced flood peaks during July–September, alongside persistent dry-season low flows, highlighting the basin’s growing exposure to hydrological extremes. The long-term projections are based on extrapolation of historical discharge patterns using AI models, without explicit climate scenario forcing (e.g., CMIP6). Accordingly, the projections are interpreted as trend-based indications rather than climate-driven predictions. While Linear Regression suggests steadily increasing discharge trends, Random Forest shows limited extrapolation capability for long-term forecasting. This limitation arises because Random Forest, as a tree-based model, has weak extrapolation capability beyond the training data range. The findings demonstrate a potentially applicable hybrid framework that enhances prediction in data-scarce transboundary basins, including other Nile tributaries and river systems such as the Mekong. The proposed framework provides a scalable and potentially applicable approach to support flood early warning, reservoir operation, and climate resilience planning, contributing to improved Earth system–based water management in transboundary river basins. The framework is directly relevant to regional basin organizations and national water management authorities, supporting evidence-based decision-making for flood early warning and reservoir operation. Graphical Abstract The graphical abstract displays a hybrid flood forecasting scheme that is established in data-sparse river basins and is shown on the Upper Blue Nile Basin in the El-Deim station. The graphical abstract illustrates a sequential hybrid flow: (1) Input of GIS-derived catchment attributes, (2) HEC-HMS baseline simulation of hydrological processes, and (3) AI refinement using ETS and Random Forest to reduce forecasting error at the El-Deim station. It starts with the use of constrained observed discharge data (2000–2006) and GIS-derived catchment attributes in an area where there is a propensity to floods as well as droughts. A hydrological model (HEC-HMS) that is physically based is initially calibrated to produce a base simulation of the key hydrological processes. This default output is combined with the data-driven models of artificial intelligence, such as Linear Regression, Exponential Smoothing (ETS), and Random Forest Regression, to increase the predictive accuracy by learning nonlinear and seasonal trends. The hybrid integration approach method greatly decreases the error of forecasting, and ETS has the lowest RMSE, as compared to the physical model alone. Long-term forecasts show the agglomeration of the flood peaks in the wet season (July September) and long-run low-flow conditions in the dry season. The suggested framework helps to early warn of floods, optimize the operation of the reservoir, and plan climate resilience, and it can provide a scalable way of forecasting hydrology in data-constrained basins.

Developing a Conceptual Model Integrating Hydrological Modeling and Artificial Intelligence for Flood and Drought Forecasting in the Upper Blue Nile Basin

期刊论文
发表日期：

2026年4月20日

查看全文

O. A. S. Walid,
Hanaa Salem Marie,
Fahmy S Abdelhaleem,
SudoyKumer Ghosh,
Anjon Das,
Manal Gad

摘要

Floods and droughts in the Upper Blue Nile Basin pose persistent challenges to water security, agriculture, and transboundary water management across Ethiopia, Sudan, and Egypt. Historical records indicate that extreme flood and drought events recur frequently in the basin, with major floods recorded in several years over the past decades (e.g., 1988, 1998, and 2014), while severe drought conditions have affected a substantial proportion of years, notably during periods such as 1984, 2002, and 2015. These recurring hydrological extremes underscore the severity of climate variability in the basin and highlight the urgent need for reliable forecasting frameworks. This study develops a hybrid hydrological–artificial intelligence (AI) framework in which a physically based hydrological model (HEC-HMS) is first used to simulate river discharge, and its outputs are subsequently post-processed using data-driven AI forecasting techniques to improve discharge prediction under data-limited conditions. The methodology combines GIS-derived catchment characteristics, observed discharge records (2000–2006), and three AI approaches Linear Regression, Exponential Smoothing (ETS), and Random Forest Regression applied at the El-Deim gauging station. HEC-HMS was first calibrated across thirteen sub-catchments between Lake Tana and El-Deim to establish a physically grounded baseline simulation. AI models were subsequently trained and validated to capture nonlinear behavior and seasonal variability in river discharge. The 2000–2006 calibration period was chosen due to limited availability of continuous discharge data. The selected AI models represent increasing complexity: Linear Regression as a baseline, ETS for capturing seasonality in short time series, and Random Forest for nonlinear relationships, while avoiding the high data demands of deep learning models such as LSTM or ANN. Results demonstrate that the hybrid framework significantly imshows forecasting performance. A paired t-test on RMSE values confirmed that the observed error reduction is statistically significant. Among the tested models, ETS achieved the highest accuracy, reducing the root mean square error (RMSE) from approximately from approximately 4 m³ s^− 1 for HEC-HMS to about 1.4 m³ s^− 1. Calibration performance was evaluated using standard statistical metrics, including the Nash–Sutcliffe Efficiency (NSE) and the coefficient of determination (R²), which indicated good agreement between simulated and observed discharges across the thirteen sub-catchments. Long-term projections (2018–2030) indicate increasing wet-season flows with pronounced flood peaks during July–September, alongside persistent dry-season low flows, highlighting the basin’s growing exposure to hydrological extremes. The long-term projections are based on extrapolation of historical discharge patterns using AI models, without explicit climate scenario forcing (e.g., CMIP6). Accordingly, the projections are interpreted as trend-based indications rather than climate-driven predictions. While Linear Regression suggests steadily increasing discharge trends, Random Forest shows limited extrapolation capability for long-term forecasting. This limitation arises because Random Forest, as a tree-based model, has weak extrapolation capability beyond the training data range. The findings demonstrate a potentially applicable hybrid framework that enhances prediction in data-scarce transboundary basins, including other Nile tributaries and river systems such as the Mekong. The proposed framework provides a scalable and potentially applicable approach to support flood early warning, reservoir operation, and climate resilience planning, contributing to improved Earth system–based water management in transboundary river basins. The framework is directly relevant to regional basin organizations and national water management authorities, supporting evidence-based decision-making for flood early warning and reservoir operation.

Graphical Abstract

The graphical abstract displays a hybrid flood forecasting scheme that is established in data-sparse river basins and is shown on the Upper Blue Nile Basin in the El-Deim station. The graphical abstract illustrates a sequential hybrid flow: (1) Input of GIS-derived catchment attributes, (2) HEC-HMS baseline simulation of hydrological processes, and (3) AI refinement using ETS and Random Forest to reduce forecasting error at the El-Deim station. It starts with the use of constrained observed discharge data (2000–2006) and GIS-derived catchment attributes in an area where there is a propensity to floods as well as droughts. A hydrological model (HEC-HMS) that is physically based is initially calibrated to produce a base simulation of the key hydrological processes. This default output is combined with the data-driven models of artificial intelligence, such as Linear Regression, Exponential Smoothing (ETS), and Random Forest Regression, to increase the predictive accuracy by learning nonlinear and seasonal trends. The hybrid integration approach method greatly decreases the error of forecasting, and ETS has the lowest RMSE, as compared to the physical model alone. Long-term forecasts show the agglomeration of the flood peaks in the wet season (July September) and long-run low-flow conditions in the dry season. The suggested framework helps to early warn of floods, optimize the operation of the reservoir, and plan climate resilience, and it can provide a scalable way of forecasting hydrology in data-constrained basins.