Research on wind power prediction technology is of great significance for mitigating the effects of randomness on the output power of the power generated by the system. First of all, the sample data is formed combined with wind power monitoring system of history in the database data and environmental parameters, at the same time genetic algorithm is used to optimize the model parameters, such as kernel function type, the kernel function parameter and error warning factor. The GA-SVM model is established which improves the efficiency of the model parameter combination optimization choice and the prediction precision. At last, based on the example verification, the standard SVM method and BP neural network method are compared. Prediction results show that the proposed GA-SVM optimization model on the ultra-short-term wind power prediction has a better learning ability and generalization ability.