Nasrin Mirzaee; Amirpouya Sarraf
Abstract
River runoff forecasting in watersheds has a special place in the management and planning of water resources for the design of water facilities, water intake from rivers, consumption management and etc. In the present study, the performance of some data integration models including simple averaging, ...
Read More
River runoff forecasting in watersheds has a special place in the management and planning of water resources for the design of water facilities, water intake from rivers, consumption management and etc. In the present study, the performance of some data integration models including simple averaging, weighted averaging and integrated artificial neural network model in monthly discharge modeling has been evaluated and compared. For this purpose, monthly flow prediction in upstream basin of Jiroft Dam was examined using Artificial Neural Network (ANN) models, Adaptive Neural-Fuzzy Inference System (ANFIS), ARIMA model and Support Vector Regression (SVR) model as an individual model. Then, the individual models were trained and validated using selected predictor variables and their results were selected for use in the integration process. Large-scale climatic signals including NAO, ENSO and PDO are also used in hydrological forecasts of river flow and the performance of single and integrated models in two modes with and without considering these signals has been compared based on the evaluation of three criteria Nash-Sutcliffe (NSE), Coefficient of determination (R2) and Mean Square Error (MSE). Results of this study indicated that the integrated approach significantly increases the accuracy of predictions. In addition, large-scale climatic signals were found to improve results, especially during the test period. For example, the results of the integrated model of artificial neural network with large climatic scale signals show that this model has the best performance among the integrated models. Also, the NSE criterion has improved by 0.04 in training compared to the integrated model of artificial neural network without large-scale signals and the MSE error has been reduced by 0.001.
