Performance comparison of Artificial Intelligence models with IHACRES model in streamflow modeling of the Gamasiab River catchment

Today due to climate change and fluctuations in the intensity and duration of rainfall in most parts of the world, new approaches to streamflow modeling have an extraordinary role in managing water resources and reducing the risks of floods. Since in some catchments, it is not feasible to measure all the observed quantities required for modeling the streamflow process, it is necessary to choose a simple model that can accurately predict rainfall-runoff using minimal information. Artificial intelligence (AI) models have high efficiency, especially when the accurate estimation of processes is more important than understanding the mechanisms and relationships that create them. Therefore, the AI models and semi-conceptual IHACRES models are used for streamflow modeling and compared with each other. In this study, streamflow modeling for the Gamasiab River, located in western Iran, is presented.

For rainfall-runoff modeling, artificial intelligence (AI) models include artificial neural network (ANN) models of type multi-layer perceptron (MLP), radial basis function neural network (RBF), and long short-term memory model (LSTM) are used. In addition, to better evaluate the AI models, a specialized semi-conceptual rainfall-runoff model called IHACRES is used. The data used in this study include daily data of flow discharge, precipitation, and average temperature for 31 years (September 23, 1986 - September 22, 2017), which is a time series of delayed data and as an input signal used to models. To evaluate the performance of the models, some criteria including Nash-Sutcliffe efficiency coefficient (NSE), root mean square error (RMSE), and correlation coefficient (R) were used.

The results show better performance of ANN, RBF, and LSTM models than the IHACRES model, especially at the peak flow rate, in modeling daily streamflow for the study area. The IHACRES model has performed well in calm and medium flows but has not performed well at flow peaks. In addition, the LSTM model performed better than the other models in estimating the flow rate during the verification period, while the ANN and RBF models performed better than the other models in the calibration phase. Overall, the results indicate that the best data-based model (i.e., LSTM model) has more than twice as good performance in streamflow modeling as the semi-conceptual IHACRES model based on RMSE criteria. In general, the results showed that artificial intelligence models are valuable tools for modeling streamflow fluctuations.

The AI models including, ANN, RBF, and LSTM models, especially in estimating flow peaks, were significantly better than the IHACRES model. The IHACRES model has performed well in calm river streamflow and low and medium discharge but has not performed well at flow peaks. In general, is recommended the LSTM model for modeling the daily streamflow of the study area due to better performance. AI models can model the streamflow more accurately and provide more efficient management of water resources in different regions. In general, the results showed that various AI models are a suitable tool in streamflow modeling, and it suggested that they be more utilized in future research.