Using artificial neural networks in prediction, runoff and sediment

This study deliberated the possibility of simulating runoff and sediment produced during individual rainfall events by ANN (Artificial Neural Network) models, in a small agricultural watershed (Badjgah plain, south of I. R. Iran). A well-known Multi-Layer Feedforward network with Backpropagation algorithm and Extended-Delta-Bar-Delta learning rule was used in the research. Considering dynamic properties, in addition to each input the former input with a reasonable lag time was introduced to the model of instantaneous graphs (Type-1). Besides, based on dissimilar features and for better performance a separate model for cumulative curves (Type-2), were developed. In order to gain all possible information for model development, in each case a network was built with the whole data as main model. Afterward, every rainfall event was selected for testing data sets whilst others were used as learning phase. Results showed that, simulation of runoff hydrograph in type-1, can be done chiefly based on the actual and 20min lagged rainfall hyetograph, antecedent soil moisture and day of water year. The main model had 0.93 as R-2 with 0.92 slope and showed more than 90% efficiency for explaining total variance and matching of data. It is realized that the predictions have been strongly affected by the oscillation in the shape of the hyetographs, and because sedograph sacted differently based on inherent processes of soil erosion and of coarse imprecise aswell as narrow-inclusive samples, the simulation did not successfully accomplish, by applied network, to simulate the feature governed on the phenomenon. Hence, the other type was suggested. Model type-2 that considered cumulative values of the hyetographs instead of instantaneous rates represented the rainfall-runoff-sediment process in a more efficient manner. The value of R2 in the main model was 0.998 with 0.994 slope in which over 99% efficiency were obtained. It is moreover concluded that modeling desirable parameters directly accompany with cumulative runoff and sediment, may be used for checking predictions and suggested in a desirable condition of precise, sufficient and all-inclusive data.