Prediction of Monthly River Flow Using Hybridization of Linear Time Series Models and Bayesian network (Case Study: Bakhtiari River)

One of the most important issues in water resources management is the preparation and development of appropriate models in order to predict the streamflow more accurately. For this purpose, in the present study, linear time series models (ARMA), intelligent Bayesian network (BN) and BN-ARMA hybrid model have been developed for forecasting the monthly river flow of Bakhtiari River in 1955-2016. The performance of the developed models was evaluated based on statistical indices such as root mean square error (RMSE), correlation coefficient (CC) and Kling-Gupta index (KGE). Among the time series models fitted to the data, the AR (3) model was selected as the appropriate model for monthly stream flow series, with the lowest value of the modified Akaike information criterion equal to 1089.3. The results showed that the AR (3) model with an error of 47.786 (m3/s) has acceptable performance. The monthly river flow from the previous month, two months and five months ago was used to model the monthly river flow using the BN model. The results indicated that with three months intervals, the model performance is optimized and its performance is weakened by increasing the number of inputs. The correlation coefficient, root mean square error and KGE in the best case of BN model in the test stage are 0.826, 45.303 and 0.789 (m3/s), respectively. Next, the combination of BN and ARMA(3.0) models was performed. The results showed that the BN-ARMA hybrid model significantly increases the accuracy of the modeling and reduces the prediction error from 45.303 (m3/s) to 15.021 (m3/s).