In this paper a class of nonlinear adaptive multi-step-prediction algorithm base d on the manifold theory was proposed. We have performed the multi-step-predicti on by exploiting images of P-step iterations of several nearest neighbors with t his method. The simulation indicated that this method was available and could im prove the prediction speed, and that the series of the standard deviation of err or after prediction has an exponential growth ratio that is the largest Lyapunov exponent.