An innovative method for dynamic update of initial water table in XXT model based on neural network technique

The initial subsurface flow of whole basin plays a quite important role in daily rainfall–runoff simulation. However, general physically based rainfall–runoff model, such as the XXT model (a hybrid model of TOPographic MODEL and the Xinanjiang model), is difficult to catch the non-linear factors and take full advantages of previous information of rainfall and runoff that is essential to the initial watershed average saturation deficit of each time step. In order to address the issue, this study selected the initial subsurface flow for the whole time series of the XXT model as the breakthrough point, and used the observed runoff and rainfall data of two days before the present day as the inputs of artificial neural network (ANN) and initial subsurface flow of the present day as the output, then integrated ANN into runoff generation module of XXT model and finally tested the integrated model for daily runoff simulation in large-scale and semi-arid Linyi watershed, eastern China. In addition, this work employ particle swarm optimization (PSO) algorithm to seek the best combination of 6 physical parameters in XXT and a great number of weights in ANN to avoid the local optimization. The results show that the integrated model performs much better than XXT in terms of Nash–Sutcliffe efficiency coefficient (NE) and root mean square error (RMSE). Hence, the new integrating approach proposed here is promising for daily rainfall–runoff modeling and can be easily extended to other process-based models.