Traditional time-domain simulation method has long calculation time on transient voltage stability margin prediction problem, the direct method is difficult to adapt to the complex power grid model, and the deep neural network method requires a large data set. In this paper, we propose the method based on LightGBM, which has the advantages of appropriate model complexity and less training time demand. Based on a regional subsystem of Guangxi Power Grid with more electrolytic aluminum load, and a power system transient voltage stability margin prediction model based on LightGBM is constructed. To solve the long tail distribution learning problem, target clipping and log transformation techniques are proposed to enhance the model. The effectiveness of them is verified by experiment results. And we propose a new metric, partial RMSE, to evaluate the performance of the model fitting the instability data. The experiment results also show that it has achieved better performance and training time than other machine learning models.