Stream flow forecasting by artificial neural network (ANN) model trained by real coded genetic algorithm (GA)

Runoff analyses are important to efficient l y manage the watersheds such as precise prediction of discharge. Runoff is mainly composed of surface and groundwater flow components, therefore hydrological conditions should be well described before any runoff analysis. In this study, runoff analyses were performed for two small sub-basins of a mountainous catchment of Tono area Japan with the aim to forecast runoff after 1 and 1/2 hours by using 3 different numerical models and performances of these models were compared to each other. The runoff and other meteorological data have been collected in these sub-basins over the last 14 years. The effect of the basin area on the prediction time of runoff and the seasonal data impacts were also investigated. For the analyses, a new approach of training artificial neural network model (ANN) with real coded genetic algorithm (GA) named as GAANN model is proposed. The results of this model were compared with famous back propagation artificial neural network (BPANN) model and with multivariate autoregressive moving average model (MARMA). It was found that for very small catchments seasonal effect on the runoff is dominant and this effect should be considered for obtaining better forecasting estimates. It was also found that estimation by ANN models was better than MARMA model for analyzing the responses to intense rainfalls in summer, whereas the results were almost similar for the light rains of winter season. The accuracy of the forecasts after several time periods in future was also investigated and found to decrease as the time period is increased. The results showed that GAANN and BPANN models almost provided similar prediction estimates in a very small mountainous watershed when precisely measured dataset was used. Modelling advantages of using genetic algorithm instead of back propagation for the training of ANN models are also highlighted.