Aiming at the problem that the accurate model parameters are not available for establishing a nonlinear dynamic system model to the data with random noise, a composite identification scheme with data preprocessing based on weighted Least Square-Support Vector Machine (LS-SVM) is proposed.According to the distribution information of the data, and by use of the robustness of weighted LS-SVM to abnormal data, the identification scheme removes anomalism of the data through regression calculation.Furthermore, the preprocessed data is used for data training to compensate for parameters of the fuzzy neural network, and the system model is obtained.The proposed composite identification scheme behaves better in the simulation of system identification with random noise.