Document Type : Research Paper
Authors
1 Associate Professor in Engineering Hydrology, Department of Rangeland and Watershed Management, Faculty of Agriculture and Natural Resources, Gonbad Kavous University, Iran
2 PhD student in Computer Engineering, Gorgan Azad University, Iran
Abstract
Introduction
Water resources management in arid and semi-arid regions is one of the fundamental challenges of the present century. Population growth, climate change, and increasing demand have put significant pressure on surface water resources. In this context, monthly runoff prediction serves as a strategic tool for reservoir planning, flood control, and sustainable watershed management. However, the complexity of hydrological processes and the nonlinear relationships among climatic variables make such predictions challenging. In recent years, hybrid models have emerged as a novel approach in hydrology that enhances prediction accuracy and stability by integrating the strengths of different models. The present study aims to evaluate the performance of linear and deep learning models and to introduce an optimal model for enhancing water resources management.
Materials and methods
In this study, monthly runoff in the Qarasu watershed, with an area of 1,624 km2, was modeled and predicted using data from three hydrometric stations (Pole-Touskestan, Naharkhoran, and Siyahab) over a common 36-year period (1985-2021). Three single models (SARIMA, BiLSTM, and GRU) and two hybrid models (BiLSTM-GRU and SARIMA-BiLSTM-GRU) were employed to model monthly runoff and forecast values for a 12-month horizon. In hybrid models, the SARIMA model is used to model the linear components of the time series, while BiLSTM and GRU models have the ability to identify and represent complex, non-linear patterns. Model performance was evaluated using RMSE, MAD, and MSE indices. The models were implemented using Python and commonly used libraries, including TensorFlow, Keras, numpy, pandas, matplotlib, scipy, and sklearn.
Results and discussion
The hybrid SARIMA-BiLSTM-GRU model, by integrating linear and nonlinear components, provided the most accurate monthly runoff predictions. The RMSE values of this model at the Pole-Touskestan, Naharkhoran, and Siyahab stations were estimated at 0.0295, 0.0173, and 0.1683 m³/s, respectively. The BiLSTM-GRU model ranked second, with RMSE values of 0.0326, 0.0226, and 0.3013 m³/s. Among the individual models, BiLSTM and GRU produced similar and relatively accurate results, while the linear SARIMA model, with RMSE values of 0.0851, 0.0230, and 0.3892 m³/s, showed the lowest performance. On average and based on the RMSE index, the SARIMA-BiLSTM-GRU hybrid model reduced prediction errors by 39.66% to 56.75% compared to the other models. The findings of this research confirm and reinforce the theoretical foundations regarding the superiority of complex hybrid models over simple, linear models.
Conclusions
This study demonstrated that hybrid models combining linear and deep learning approaches can significantly improve the accuracy and stability of monthly runoff predictions. The modeling and validation results indicated that the hybrid SARIMA-BiLSTM-GRU model, which effectively combines linear (SARIMA) and nonlinear (BiLSTM and GRU) components, was selected as the optimal model for this watershed, demonstrating the best performance in monthly runoff prediction. The results highlight the importance of applying hybrid approaches in water resources management and flood control. It is recommended to employ more advanced models and diverse input variables, such as precipitation and temperature, to further enhance the accuracy and reliability of predictions. Furthermore, given the observation of overfitting in the base LSTM and GRU models, it is recommended to employ more advanced architectures, such as models based on the Attention mechanism like the Transformer, to enable the model to intelligently focus on the key and influential time periods in the runoff process.
Keywords
)