A new method to determine delay time is proposed, which is called nonlinear multiple autocorrelation function method. The method uses a high order multiple autocorrelation function, R(τ), to calculate the nonlinear correlation of system and estimates the optimal delay time by finding the first local minimum value of R(τ). The time complexity is low and the dependence on the length of data is not strong. The performance of the algorithm is tested by being applied to five noisy chaotic time series. The time series are generated by four chaotic maps adding different noise level Gauss white noise. The numerical results show that, the method is more appropriate and robust to noise.