It can be used to find out transformer faults through the predict detection of transformer oil dissolved gas. By multivariate time series reconstruction of the state variables as LSSVR model inputs, a transformer fault prediction model was proposed. Firstly, the prediction based on multivariate reconstruction principle and LSSVR theory were given. Then, it was discussed the impact of the reconstruction parameters and LSSVR parameters for predicting errors. Genetic algorithm was adopted to ensure prediction accuracy. Finally, the method was used to verify the applicability of support vector machine prediction based on Lorenz system. Compared with other predictive approaches, the proposed combinational forecast model has higher prediction accuracy.