Due to the fact that linear prediction method is difficult to describe the nonlinear, non-stationary changes of cloud clusters, a technique of retrieval nonlinear clouds clusters forecast model, based on the idea of combining the decomposition of empirical orthogonal function (EOF) and the genetic algorithm optimization parameters, was presented. Firstly, satellite image sequences were temporal-spatially decomposed by EOF. On this basis, genetic algorithms were introduced to make the dynamic model reconstruction and model parameters optimization retrieval of EOF time coefficients sequence, and a nonlinear differential equations of EOF time coefficients were established. Then, by the EOF temporal-spatial functions synthesis, a dynamic forecast model of cloud clusters evolution was structured. The experimental results showed that the retrieved clouds dynamic forecast model was more reasonable in describing the cloud evolution of the underlying trend in particular seasons and region, and the forecast results were better accorded with the basic characteristics of actual satellite cloud pictures. Especially, a middle-long period over three hours objective cloud clusters predictions was achieved.