University of TehranIranian Journal of Soil and Water Research2008-479X51820201022Application of NARX Neural Network as Surrogate Model to Long-term Simulation of the Outlet Salinity from Strong Stratified ReservoirsApplication of NARX Neural Network as Surrogate Model to Long-term Simulation of the Outlet Salinity from Strong Stratified Reservoirs197119827663610.22059/ijswr.2020.299151.668532FAMoradAsadiWater Structures Engineering Department, Tarbiat Modares University, Tehran, IranJamalMohamad Vali SamaniDepartment of Civil Engineering, Shahr-e-Qods Branch, Islamic Azad University, Tehran, IranHosseinMohamad Vali SamaniDepartment of Civil Engineering, Faculty of Engineering, Shahid Chamran University, Ahwaz, IranJournal Article20200314The CE-QUAL-W2 program as a physical model for quality and hydrodynamic simulation of water reservoirs has a high computational cost. Therefore, finding surrogate models to give optimal results in short term would have a great practical importance especially in simulation-optimization problems. In this study, the capability of the NARX model as a surrogate model was investigated to simulate the outlet salinity from strongly stratified reservoirs. For this purpose, the CE-QUAL-W2 model was used and calibrated to simulate the outlet salinity of the Upper Gotvand Reservoir over 10 years. Regarding the possibility of release from different reservoir intakes, by monthly change of release ratios, several problems were defined and a library of the physical model results was formed. Then different NARX architecture scenarios were introduced and trained using the library results. The results obtained from different scenarios indicate that the NARX neural network model has a high capability to simulate the CE-QUAL-W2 model results of outflow salinity, so that the correlation coefficient is always above 0.91. In the selected scenario, a very good agreement is observed between the results of the two models, with a correlation coefficient of 0.95, mean absolute percentage error of 8.7% and Nash-Sutcliffe coefficient of 0.79. The simulation time required for the NARX neural network model is less than 0.06% of the time required to run the physical model for the same problem. The results show that the NARX model can be used as a suitable surrogate model for CE-QUAL-W2 to predict the long-term reservoir outlet salinity and reduces the cost of computing while maintains accuracy.The CE-QUAL-W2 program as a physical model for quality and hydrodynamic simulation of water reservoirs has a high computational cost. Therefore, finding surrogate models to give optimal results in short term would have a great practical importance especially in simulation-optimization problems. In this study, the capability of the NARX model as a surrogate model was investigated to simulate the outlet salinity from strongly stratified reservoirs. For this purpose, the CE-QUAL-W2 model was used and calibrated to simulate the outlet salinity of the Upper Gotvand Reservoir over 10 years. Regarding the possibility of release from different reservoir intakes, by monthly change of release ratios, several problems were defined and a library of the physical model results was formed. Then different NARX architecture scenarios were introduced and trained using the library results. The results obtained from different scenarios indicate that the NARX neural network model has a high capability to simulate the CE-QUAL-W2 model results of outflow salinity, so that the correlation coefficient is always above 0.91. In the selected scenario, a very good agreement is observed between the results of the two models, with a correlation coefficient of 0.95, mean absolute percentage error of 8.7% and Nash-Sutcliffe coefficient of 0.79. The simulation time required for the NARX neural network model is less than 0.06% of the time required to run the physical model for the same problem. The results show that the NARX model can be used as a suitable surrogate model for CE-QUAL-W2 to predict the long-term reservoir outlet salinity and reduces the cost of computing while maintains accuracy.https://ijswr.ut.ac.ir/article_76636_ca8b72d8e64f4e89ad2248c98e0c4021.pdf