Delay of flights is a core problem in civil aviation industry. It is significant to predict flight delay situation in advance so as to take reasonable measures to negative effects of flight delay. In order to improve the prediction performance， a flight delay multi-classification prediction model based on light gradient boosting machine （LightGBM） is put forward in this paper. This model can screen features by using variance filtering and recursive feature elimination according to flight information and weather information. It uses synthetic minority oversampling technique （SMOTE） and Tomek Link to deal with unbalanced data. Finally， LightGBM is used to build models， and multi-classification prediction of flight delay lengths can be realized. In order to verify the rationality of the model， this paper compares the proposed model with models constructed by other advanced algorithms. The experimental results show that the proposed model performs better in terms of various prediction performance indexes and can improve the prediction accuracy to 90% or higher. The model can also greatly reduce the training time and cost.