Extended Abstract
Introduction
Population growth and the limited availability of surface and groundwater resources have made the optimal management and utilization of water resources an essential necessity. In this regard, investigating the trends of temperature, precipitation, and streamflow variations can play a significant role in water resources management. Temperature, precipitation, and discharge are among the fundamental elements for understanding the climate and hydrology of a watershed. Predicting hydrological parameters has received considerable attention due to their importance in drought analysis and water supply, the design of hydraulic structures, reservoir operation, erosion and sediment control, and many other water-related applications. Rivers are not only considered direct sources of water supply for consumers, but they also play a vital role in maintaining ecosystem stability, recharging groundwater resources, and supporting biodiversity. However, extensive climate changes in recent decades, including rising temperatures, altered precipitation patterns, recurrent droughts, and sudden intense rainfall events, have directly and indirectly affected river flow regimes. Owing to the importance of this issue, numerous studies have been conducted worldwide. Therefore, the present study focuses on river flow modeling with an emphasis on low-flow and high-flow periods using several deep learning models (LSTM and GRU) and machine learning models (RF and MLR) at selected stations within the Karkheh watershed.
Materials and Methods
From a geomorphological perspective, the study area is located in the central part of the Zagros Mountain range in Lorestan Province, Iran, between the geographical longitudes of 47°12′30″ to 48°59′20″ E and latitudes of 33°05′45″ to 34°03′26″ N. This study was conducted to predict river discharge under low-flow and high-flow conditions using RF, MLP, LSTM, and GRU models. In the first stage, the required data were collected. The datasets included monthly river discharge, precipitation, temperature, and evaporation records over a 10-year period (2014–2015 to 2023–2024) obtained from the Dareh Tang Kahman, Afrineh Cholehhol, and Afrineh Kashkan hydrometric stations. These data were categorized into two separate groups representing wet and dry conditions. Subsequently, the data preprocessing stage was carried out. After normalizing the datasets, the data were divided into training and testing subsets, with 70% allocated for model training and 30% for model testing. Temperature, precipitation, and evaporation were considered as the input variables, while river discharge was defined as the output variable of the models. Two deep learning models with similar architectures, consisting of recurrent layers and an output layer, were designed and trained. Each model was implemented separately for wet and dry periods using the Python programming language. Finally, the performance of the models was evaluated and compared using statistical error criteria, including the correlation coefficient (R), mean absolute error (MAE), and root mean square error (RMSE). In addition, sensitivity analysis of the models with respect to different input variables was performed using the elimination method in order to identify the most influential input parameter affecting the model outputs.
Results and Discussion
The results demonstrated that, based on the evaluation criteria, the GRU model during the high-flow period and the RF model during the low-flow period outperformed the other models. According to the testing results, the GRU model provided the best performance during the high-flow period. The correlation coefficient (R), mean absolute error (MAE), and root mean square error (RMSE) values obtained for the Afrineh Cholehhol station were 0.95, 0.06, and 0.10, respectively. Similarly, for the Afrineh Kashkan station, these values were 0.57, 0.17, and 0.24, respectively, while for the Dareh Tang station, the corresponding values were 0.94, 0.09, and 0.13. In addition, the testing results indicated that the RF model achieved the best performance during the low-flow period. The values of R, MAE, and RMSE for the Afrineh Cholehhol station were 0.72, 0.22, and 0.23, respectively. For the Afrineh Kashkan station, the obtained values were 0.82, 0.17, and 0.20, respectively, and for the Dareh Tang station, they were 0.43, 0.34, and 0.38, respectively. The sensitivity analysis of the best-performing models in both high-flow and low-flow periods revealed that precipitation was the most influential parameter affecting the model outputs. The removal of this parameter resulted in a noticeable decline in model performance, indicating its critical role in accurate river discharge prediction.
Conclusion
Overall, the modeling results demonstrated that both deep learning and machine learning models performed well in simulating river discharge during the studied periods using the selected input data at the investigated stations. The findings also indicate that these models have the capability to be applied to neighboring stations with similar hydrological conditions. Furthermore, the results revealed that the considered input parameters, particularly precipitation, had a significant influence on river discharge, leading to high model accuracy in discharge simulation and prediction. Therefore, based on the findings of this study, these optimized models can be effectively utilized to reduce both time and costs in water and soil conservation studies as well as in estimating watershed outlet discharge. In addition, for improved management of the quantity and quality of surface water resources, these models can be applied to estimate discharge at nearby ungauged stations with similar geological and hydrological conditions within the region, providing reliable results regarding river discharge. |
