By using non-parametric kernel density estimation method, a NP stochastic simulation model of Urumqi River monthly runoff was established. Moreover, a particle swarm optimization model using a LSCV (Least Squares Cross - Validation) as the objective function was employed to obtain bandwidth parameters; a variable kernel bandwidth method was adopted to adjust kernel density boundary. Afterwards, 53-year (1958-2010) records of monthly runoff were applied for 250 simulations, each with a length of 53 years, were made to carry on the practicability test of the model. Finally, the results from a PAR (seasonal autoregressive) model built by software SAMS2007 was presented for comparison of NP model. The results show that the Urumqi River monthly runoff NP model can better maintain the statistical properties of the original sequence. Comparing with the PAR model, NP model has the characteristics of fewer parameters and simple calculation.