Abstract: In order to improve the flow concentration accuracy of the Xinanjiang model and to reduce the influence of personal experiences on the model calibration, a new rainfall-runoff model called XBK (XAJ-BP-KNN) is developed, coupling the Xinanjiang runoff generation model with the improved version of the back propagation (BP) flow concentration model. The latter uses the BP neural network algorithm to simulate the nonlinear relationship of the flow concentration process. The flow calculated by Xinanjiang runoff model and antecedent flow are used as the XBK inputs to a BP simulation network. The flow inputs are routed by the BP concentration model to the outlet of the network, which forms the hydrograph at the outlet of the BP simulation network. XBK uses the similarity theory and the K-nearest neighbor algorithm for pattern recognition in an effort to correct the simulation error due to the absence of the observed initial flow data. XBK parameters are optimized globally using the combined method of the shuffled complex evolution (SCE-UA) algorithm and the genetic early stopping Levenberg-Marquardt (LM) algorithm. The XBK model is applied to the Chengcun watershed. Compared to the original version of the Xinanjiang model, the result shows that a better model simulation can be achieved with XBK. XBK is easy to apply, and the combined global optimization algorithm is able to identify optimal parameter values.