Evaluation of the ability of adaptive neuro-fuzzy interface system, artificialneural network and regression to regional flood analysis

Developing of techniques for regional flood frequency estimation in ungauged sites is one of the foremost goals of contemporary hydrology. The flood frequency evaluation for ungauged catchments is usually approached by deriving suitable statistical relationships (models) between flood statistics and basins characteristics. Already, several equations have been presented to estimate the flood frequency in different areas such as Karkheh basin. However, due to the complexity of this phenomenon, the relationships have not been capable to simulate the flood frequency with desired accuracy. Accordingly, in this study, in addition to the regression method has been used in the previous studies, the ANN and ANFIS models are applied. In fact, these are a type of black box models without any knowledge of processes within the system, in which inputs are converted into outputs (or output). This situation indicates that this type of new models is actually similar to the regression relations, however, there is further flexibility in adjusting the weights and thus can be used as an replacement to multivariate regressions.
The study area, including 33 hydrometry stations, is located in the west of Iran. In this study, 27 of the stations for calibration and 6 of the stations for validation were used. To approach a unique model, return period was taken into account as the independent factor.
For achieving the best ANN and ANFIS system, different combinations of physiographic with return periods, as input data, has been used. To find the important input factors of the models, sensitivity analysis has been performed in SPSS software. Accordingly, the most important independent variables were including: Return period, area, height, main stream length and slope. In the ANN model, different combinations of these inputs were compared together. It should be noted that for optimizing the connecting weights among different layers of ANN, Genetic algorithms have been used. Consequently, the best selected network is Feed-forward with the structure of 5-10-1 and R=0.95. In the ANFIS system, with increasing the number of input variables for each of the four membership function, including Triangular, Gaussian, Gaussian2 and trapezoidal, simulation accuracy increases. The best simulation is a triangular function with RMSE=0.1514, R=0.97 and the number of rules is 243. Finally, by comparing models, The ANFIS model was selected as the best model. The ANFIS has the best accuracy especially in high return period.
Where the sub-basins are small and their flood in different return periods is less than 1000 m3/s, the regression model makes a good accordance with real flood. The ANN model has also good performance in low discharges. The regression presents its forecast in the framework of formulas and it is better and more practical for engineers. Generally, The ANFIS model is the best model for all ranges of the discharge and the best tool for prediction enormous flood in Karkheh basin.