This paper explores all student related data using educational data mining techniques to draw conclusions about their performance and behavior. The main idea and theoretical basis of Random Forest is described, the importance of each feature is calculated using Random Forest based Important Feature Selection Algorithm, and each feature is ranked and the best feature is selected as the effective feature for constructing the performance prediction model. By Light GBM is to further improve the GBDT algorithm and XGBoost algorithm to construct the Light GBM grade prediction model in order to improve the training speed and the prediction ability of the model. In order to verify the feasibility of the constructed model, the application of sustainable education in Chinese language education is tested from various aspects such as model testing as well as learning behavior. The results show that the accuracy of the Ligth GBM grade prediction model increases from 0.68 to 0.918 when the course progress is raised from 10 to 100, i.e., the accuracy of the Ligth GBM grade prediction model in predicting students’ grades gradually increases as the course progresses, so that it can effectively analyze the application of sustainable education in Chinese education.