Integrated neural networks for monthly river flow estimation in arid inland basin of Northwest China

Streamflow model including rainfall–runoff and river flow models play an important role in water resources management, especially in arid inland area. Traditional conceptual models have the disadvantage of requirement of spatial variation parameters about the physical characteristics of the catchments. To overcome this difficulty, in this study, several integrated Artificial Neural Networks (ANNs) were presented to estimate monthly river flow, and the models include the semi-distributed forms of ANNs that can explore spatial variation in hydrological process (such as rainfall distribution and evaporation distribution) and no requirement of physical characteristic parameters of the catchments. In an arid inland basin of Northwest, integrated ANNs were developed using hydrological and agricultural data, and its performance was compared with that of lumped ANN and local linear regression model (LLR). Results showed that the integrated ANNs perform well to estimate the monthly streamflow at outlet of mountain with Root Mean Square Error (RMSE) of 0.36 × 107 m3 and Relative Error (RE) of 9%. Similarly, the integrated ANNs can also accurately estimate the monthly river flow downstream of the basin with RMSE of 0.35–0.38 × 107 m3 and RE of 22–27%. When compared with integrated ANNs, the lumped ANN and LLR models have lower precision to simulate monthly streamflow in arid inland basin. Presented integrated ANN models retain the advantages of the semi-distributed models considering the heterogeneity and spatial variation of hydrological factors and the physical characteristics in the catchment, while taking advantage of the potential of ANNs as an effective tool in nonlinear mapping or functional relationship establishment. In contrast to traditional models either in the lumped ANN or in empirical regression forms, the new approach of integration of Artificial Neural Networks has shown great potential in streamflow modeling.