Open Access

ARTICLE

Innovative Deep Learning Models for Streamflow Forecasting in High Elevation Catchments

Rana Muhammad Adnan Ikram¹, Jing-Cheng Han^1,*, Ahmed A. Ewees², Mo Wang³, Ozgur Kisi^4,5,6,*, Salim Heddam⁷, Mohammad Zounemat-Kermani⁸

1 Water Science and Environmental Research Centre, College of Chemistry and Environmental Engineering, Shenzhen University, Shenzhen, China
2 Department of Computer, Damietta University, Damietta, Egypt
3 College of Architecture and Urban Planning, Guangzhou University, Guangzhou, China
4 Department of Civil Engineering, Lübeck University of Applied Science, Lübeck, Germany
5 Department of Civil Engineering, School of Technology, Ilia State University, Tbilisi, Georgia
6 School of Civil, Environmental and Architectural Engineering, Korea University, Seoul, Republic of Korea
7 Faculty of Science, Agronomy Department, Hydraulics Division University, 20 Août 1955, Route El Hadaik, BP 26, Skikda, Algeria
8 Department of Civil Engineering, Shahid Bahonar University of Kerman, Kerman, Iran

* Corresponding Authors: Jing-Cheng Han. Email: ; Ozgur Kisi. Email:

(This article belongs to the Special Issue: Explainable AI, Digital Twin, and Hybrid Deep Learning Approaches for Urban–Regional Hydrology, Water Quality, and Risk Modeling under Uncertainty)

Computer Modeling in Engineering & Sciences 2026, 148(1), 31 https://doi.org/10.32604/cmes.2026.083313

Received 01 April 2026; Accepted 05 June 2026; Issue published 27 July 2026

Abstract

Two-phase optimized machine learning and deep learning models play a key role in enhancing the prediction accuracy of nonlinear time series modeling. This study assesses the performance of a novel two-phase optimized Long Short-Term Memory (LSTM) model with integration of Aquila Optimizer (AO) and Wild Horse Optimizer (WHO) in predicting monthly streamflow in a snow-fed catchment. The two-phase optimized LSTM-WHOAO model is compared with single-phase optimized models such as LSTM-GA (Genetic Algorithm), LSTM-GWO (Grey Wolf Optimizer), LSTM-WOA (Whale Optimization Algorithm), LSTM-AO, and LSTM-WHO. The outcomes acquired from the deep learning models were compared using four statistical measures: root-mean-square-error (RMSE), mean absolute error (MAE), Nash-Sutcliffe efficiency (NSE), and coefficient of determination (R²). The LSTM-WHOAO model exhibited the best performance during training, with a mean RMSE of 51.930 and an R² of 0.851. The LSTM-WHO model also demonstrated robust performance, achieving an average RMSE of 53.900 and an R² value of 0.840. Other models, such as LSTM-AO and LSTM-WOA, showed average RMSE of 57.135 and 58.978, respectively, indicating improved performance over the single LSTM model. For the testing stage, the LSTM-WHOAO model remained more effective than other models, with an average RMSE of 70.413 and an R² of 0.755. For peak streamflow events, the LSTM-WHOAO model had the lowest absolute error (201.4%), significantly reducing prediction error compared to other models such as LSTM-GA (333.9%) and LSTM (347.1%). For peak streamflow events, the LSTM-WHOAO model had the lowest absolute error (201.4%), significantly reducing the prediction error compared to other models such as LSTM-GA (333.9%) and LSTM (347.1%). Models that incorporated snow-covered area (SCA) data, such as LSTM-WHOAO and LSTM-WHO, showed lower RMSE and higher R² values, underscoring the importance of considering snow cover dynamics in streamflow forecasting. The LSTM-WHOAO model proved to be the most successful, showing superior results in both the training and testing stages, as well as in peak streamflow predictions. By addressing the unique challenges of snow-fed catchments, this research offers valuable insights, especially into the application of advanced ML techniques in hydrology.

---

Keywords

---