RAINFALL-RUNOFF MODELING OF A WATERSHED USING ANN

Abstract

Rainfall-runoff modeling is one of the most complex areas of research in hydrology because of the uncertainty of hydrological, geological and meteorological parameters and the scarcity of adequate records. Accurate modeling of rainfall-runoff relationship is important for water management and planning of available water resources. Artificial neural networks (ANN) can be an efficient way of flow modeling and is capable of modelling nonlinear complex systems and the modeller does not require any physical law that underlie the process. It has a flexible mathematical structure which is capable of identifying the non-linear relationship between input and output data sets. Radial basis function artificial neural network (RBFANN) is a type of ANN recently used by many scientists in hydrological modeling. The RBFANN network model is motivated by the locally tuned response. Owing to this ability, the networks are easily trained by using a sufficiently large data set. The study deals with the application of RBFANN network in rainfall-runoff modeling for Vamsadhara river basin of India with an available daily rainfall-runoff data. The training of RBFANN network can be split into an unsupervised part and a supervised part. Unsupervised training techniques are relatively fast. Clustering algorithm k-means is used for unsupervised learning in function layer. Gradient descent algorithm is used for supervised learning part in output layer. The unsupervised learning algorithm is separated into static and dynamic based on the method of estimation of spread value. In static models, a constant spread value is used where as in dynamic models the spread value changes in all successive iteration. The performance of both static and dynamic model critically depends on learning rates, number of iteration and network architecture. Both static and dynamic model behaves differently with respect to function layer, learning rate and number of iteration required to obtain improved model performance. Further, a fixed learning rate (ALR) is used in function layer and varying learning rate (ALRG) is used in output layer to optimize the model performance. The static RBFANN model needs less iteration compared to dynamic model in optimization. A program was developed in FORTRAN environment with a flexibility to change the input and output variables as well as to change the radial basis nodes. The program can be readily used for the modeling and simulation of hydrological applications.