What question did this study set out to answer?

This research aims to enhance streamflow prediction accuracy in data-scarce regions using machine learning and deep learning methods.

April 28, 2026Open Access

Stream flow prediction utilizing deep learning models in the Lake Abaya-Chamo sub-basin, South Ethiopia

Key Points

This research aims to enhance streamflow prediction accuracy in data-scarce regions using machine learning and deep learning methods.
Utilized deep learning models including LSTM, BiLSTM, GRU, and Conv1D-LSTM for streamflow and rainfall data reconstruction.
Reconstructed rainfall records at Arba Minch and Dilla stations and filled streamflow gaps at Kulfo and Gidabo Rivers.
Developed and tested six DL architectures and one process-based model using various hydro-meteorological inputs.
The ensemble model achieved NSE values of 0.97 (training) and 0.95 (testing) at Kulfo, and 0.97 and 0.96 at Gidabo.
NSE values for the ensemble model in flow categories were 0.65 (low), 0.73 (medium), and 0.96 (high) at Kulfo, and 0.45, 0.56, and 0.95 at Gidabo, respectively.
The HBV model performed the lowest in capturing nonlinear patterns, highlighting the superior performance of ensemble methods.

Abstract

Abstract Reliable streamflow prediction is essential for managing water resources sustainably and reducing hazards. Hydrological modeling is however still restricted by data gaps and a lack of hydro-meteorological records, especially in areas with insufficient data. The present study utilizes machine learning (ML) and deep learning (DL) methodologies to rebuild lacking rainfall and streamflow data, enhancing daily streamflow modeling in the Kulfo and Gidabo watersheds of the Lake Abaya–Chamo sub-basin, Ethiopia. The Long Short-Term Memory (LSTM) model was used to reconstruct incomplete rainfall records at the Arba Minch and Dilla stations, while Support Vector Regression (SVR) and Random Forest (RF) models were used to fill in streamflow gaps at the Kulfo and Gidabo Rivers, respectively. The study developed and tested six DL architectures and one process based models utilizing various hydro-meteorological inputs through LSTM, bidirectional LSTM (BiLSTM), gaterd recurrent unit (GRU), one dimensional convolutional neural network (Conv1D)-LSTM, Hydrologiska Byråns Vattenbalansavdelning (HBV), HBV-LSTM, and an ensemble of models. The ensemble model successfully surpassed the rest, with NSE values of 0.97 (training) and 0.95 (testing) at Kulfo, as well as 0.97 and 0.96 at Gidabo, indicating higher accuracy and stability. The HBV model, on the other hand, performed the lowest owing to incapability to capture nonlinear and complex patterns. Additionally, the ensemble model showed notable predictive ability for low, medium, and high flow categories with NSE values of 0.65, 0.73, and 0.96 at Kulfo, and 0.45, 0.56, and 0.95 at Gidabo, respectively. In summary, the combination of ML and DL models improved streamflow prediction in basins with limited data and successfully restored missing hydrometeorological data. The ensemble model demonstrated a dependable performance within the study region and indicates potential for broader applicability for managing drought, forecasting floods, and designing adaptive water resources in the face of growing hydro-climatic variability and physiographic conditions.

Perguntar à IA

Bookmark

View Full Paper