Performance Comparison of Autoregressive Runoff Prediction Methods for Different River Basins

Runoff prediction for the next month with historical stream flow data is concerned by many researchers due to its importance for field practices.Therefore,many data-driven models were developed for forecasting runoff for the next month.Since lack of reliable weather forecasting for the next month and randomness and nonlinearity of monthly runoff time series,the performances of those developed runoff prediction methods present significant differences.In this paper,ARMA model,ANN model and SVR model are used to forecast runoff for the next month at three river basins.The MAPE is used to evaluate the forecasting performances of these three models.The MAPE comparison results show that the performance of the SVR model is the best one among three models.The comparison of forecasting results among three river basins demonstrates that MAPE values are significant different for three river basins,which might be caused by different coefficient of variance (CV) of historical runoff time series and the correlation coefficient (Rlag1) between two adjacent months' runoffs.Moreover,MAPE values of three river basins have significant linear correlation with both CV and |Rlag1| values.When CV and |Rlag1| are both used to calculate MAPE,the coefficient of determination is 0.80.Further analysis shows that the difference of river basin characteristics is the crucial reason resulting in the different MAPE values (forecasting performance) for three models at three river basins. © 2018, The Editorial Board of Journal of Basic Science and Engineering. All right reserved.