A new approach for simulating and forecasting the rainfall-runoff process within the next two months

Abstract

In this study, a new approach is presented to predict rainfall and runoff in Tolt River basin within the next two months. For this purpose, a combination of wavelet transform and artificial neural network (WANN) incorporating both observed and predicted time series in the input structure is taken under consideration. To simulate the runoff process in 2months ahead, the best forecasts of the runoff in 1month ahead and the rainfall in 1 and 2months ahead are added into input variables. The best predictive model developed for runoff 1month ahead in its input structure comprises of historical time series of runoff, rainfall and temperature which demonstrates the influence of the thermal characteristics in mountainous regions. All the input variables whether predicted or observed time series are decomposed via wavelet transform and are imposed to the ANN models. A comparison between performance of the proposed WANN models with those of the traditional WANN models (using only observed time series as input variables) reveals the superiority of the new models. An uncertainty analysis is implemented to evaluate the reliability of the predictions. Results of this study demonstrate that a reliable prediction of the rainfall and runoff process both for one and two months ahead can be achieved when the new methodology is applied. According to this study, adding predictions can be an efficient way to increase the model accuracy in multi time step ahead prediction of rainfall and runoff.