Abstract: In order to explore the forecasting method which is more in line with the flood control requirements of the middle and small river basin and improve the accuracy of flood forecasting,taking Tunxi Basin as an example,combined with the actual flood forecasting requirements of the middle and small rivers,a non-equal-weight parameter calibration method with the qualified rate of flood peak and the qualified rate of flood peak onset time as the main constraints is adopted(i.e.,the weights of runoff depth,flood peak discharge,qualified rate of peak onset time and deterministic coefficient in the objective function are respectively（1∶2∶2∶1）to calibrate the parameters of the Xin’anjiang River model,and use the arithmetic mean method to couple the calculation results of the Xin ’anjiang River model and BP neural network model so as to improve the accuracy of flood prediction. The results show that the method based on the main constraints of flood peak discharge and peak occurrence time is feasible in the flood prediction of Tunxi Basin. Compared with the traditional equal-weight method,the method has more advantages in the flood peak and peak occurrence time prediction,and meets the flood control requirements of small and medium-sized rivers. The Xin ’anjiang River model can simulate the flood peak and peak time well,and the BP neural network model can simulate the flood peak and runoff depth well. The arithmetic mean method is used to couple the simulation results of the two models can improve the accuracy of flood prediction.