Estimation of Streamflow Using Data Driven Models

Abstract

The collection and analysis of hydrological data are quite important in the development, planning and management of water resources. Streamflow estimation is of great importance in terms of planning of water resources, water budget at basin scale and basin modeling studies. Artificial neural networks are a widely used technique in the estimation of streamflow, as well as in the estimation of many hydraulic and hydrological parameters. In this study, it is aimed to make a monthly average flow estimation of Sarız River Şarköy station, numbered D18A032, located in the north of Seyhan Basin by using two different Artificial Neural Network techniques. For this purpose, both the flow data of station D18A032 and the monthly total precipitation data of Sarız station numbered 17840 were used. Streamflow prediction models were obtained by using flow and precipitation data for the period 1990-2017 and nine different input combinations. Artificial neural network techniques used are Multi-Layer Perceptron and Radial Basis Function Network techniques. The generated Multilayer Perceptron models are optimized in terms of two hidden layers, the number of neurons in the hidden layers, the momentum coefficient and the learning rate. Radial Basis Function Network models are optimized for the number of spreads and the number of neurons in the single hidden layer used. In both techniques, the parameters of the most successful models created for different input combinations were determined by the trial and error method. Nash Sutcliffe efficiency coefficient, coefficient of determination, combined accuracy, root mean square error and mean absolute error performance metrics were used to evaluate the models applied. The most successful streamflow estimation in Multi-Layer Perceptron models was obtained in the M6 model, in which the parameters Qt-1, Qt-2, Qt-3, Pt, Pt-1, Pt-2, Pt-3 were used as inputs. In Radial Basis Function Network models, the most successful estimation model is the M2 model in which Qt-1 and Qt-2 parameters are used as inputs. Among the two models, the most successful model was the Multilayer Perceptron model with the value of Nash Sutcliffe = 0.717.