مقایسه مدل‌های آموزش عمیق در پیش‌بینی جریان رودخانه در غرب کشور و بر روی رودخانه کشکان

Background and Objectives
In recent years, the use of artificial intelligence methods, such as artificial neural network models, have become increasingly prevalent in simulating complex natural phenomena, including daily streamflow. The streamflow directly correlates with flood occurrences, and mitigating financial and human losses due to floods is crucial. Accurate streamflow simulation is essential for water resource management and river management. Consequently, in hydrology, deep learning methods have emerged as novel tools to address the longstanding challenge of daily streamflow modeling and are widely used in simulations.
Advancements in Streamflow Modeling with Artificial Intelligence: In recent years, the field of hydrology has witnessed a significant shift toward leveraging artificial intelligence (AI) techniques for streamflow modeling. Among these methods, artificial neural network (ANN) models have gained prominence due to their ability to capture complex relationships within hydrological systems. Streamflow, which represents the flow of water in rivers and streams, is a critical variable for understanding water availability, flood risk, and ecosystem health. By accurately simulating streamflow, researchers and water resource managers can make informed decisions regarding water allocation, flood preparedness, and environmental conservation. Hydrological processes are inherently nonlinear and influenced by various factors such as precipitation, temperature, land cover, and soil properties. Traditional hydrological models often struggle to capture these complexities. However, deep learning methods, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs), offer promising solutions. These models can learn intricate patterns from historical streamflow data, adapt to changing conditions, and provide accurate predictions. As a result, they have become indispensable tools for addressing the longstanding challenge of daily streamflow modeling. Researchers continue to explore novel architectures, data augmentation techniques, and hybrid approaches to enhance the performance and robustness of AI-based streamflow simulations. In summary, the integration of deep learning methods into hydrological research has revolutionized streamflow modeling, enabling more accurate predictions and informed decision-making in water management and flood risk assessment.

Methodology
In this study, we focused on selecting an appropriate input scenario for deep learning models and simulate daily streamflow on the Kashkan River using LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) deep learning methods. Prior to this, deep learning modeling with the GRU approach using native streamflow measurements had not been performed for Kashkan river. The study area is a flood-prone and mountainous region, specifically the western part of Iran, where a hydrological station with a history of flood events is situated on the Kashkan River. We employ four approaches for handling outliers (Mahalanobis, critical interval removal, Z-Score, and no removal) and four different preprocessing techniques for input data to train two models: LSTM and GRU. Ultimately, eight distinct models are generated and validated against historical data. The input features include regional average precipitation, normalized vegetation cover index, surface soil moisture, groundwater flow, and the Kashkan River’s own flow at the hydrological station, with the best features selected using statistical correlation control.
Findings
The results demonstrate that among the deep learning models generated with a 10-day time step, the model with the least error and consistent low error retention in error metrics is observed. Furthermore, the best performance is achieved using different approaches, in the following order: the GRU model with Z-Score-corrected inputs, followed by the Mahalanobis removal approach with average RMSE (Root Mean Square Error) and KGE (Kling-Gupta Efficiency) values of 5.41 and 0.99, respectively, and the critical interval removal approach with RMSE of 6.23 and KGE of 0.7.The results showed that among the deep learning models produced with a time step of 10 days in the model, the lowest amount of error and the persistence of low error can be seen in the error statistics, and among the different approaches used, the best performance is the GRU model with input modified by Z-Score elimination of outlier method, Mahalanobis elimination method with average RMSE and KGE values of 5.41, 0.99, 6.23, and 0.7 in the training phase and 8.17, 0.79, 4.21, and 0.81 in the validation phase and 5.01, 0.68, and 7.21 and 0.52 are in the testing phase. The obtained results do not reject the LSTM method in simulating the river flow, but state that the listed scenarios, especially in the GRU method, have a higher power in dealing with the data and recognizing the complex pattern of daily river flow, taking into account the limitation in use They have seven years of regular daily data, and future research will show how the behavior of GRU and LSTM models will differ if data with higher convergence is used.
Conclusion
GRU in future studies can make difference by (Enhanced Flood Forecasting Accuracy, Efficient Computation and Real-Time Applications, Integration with Lag Time Preprocessing, Adaptability to Changing Climate and Urbanization) future studies will be on data driven method in flood prone areas. there remains ample room for future research and innovation. Here are some directions for further exploration: Hydrological Data Fusion, Spatially Explicit Models, Uncertainty Quantification, Climate Change Resilience.