Considering the effect of generator steady-state output reactive power on the ability to support transient voltage recovery stability, preventive control through optimizing that power to improve the ability of grid maintaining transient voltage stability under severe fault disturbance is studied based on machine learning. The method uses a weighted multi-binary table transient stability margin index to quantify the comprehensive stability margin of bus transient voltages to different expected fault disturbances. Then it determines the bus with weak stability margin based on the index ranking. At the same time, based on the comprehensive action sensitivity ranking of generator reactive power regulation on the voltage transient stability margin of each weak bus under the expected fault, the action sensitive generator can be determined. Then an XGBoost classification model for predicting transient voltage stability based on the system steady-state characteristics vector is developed. This considers the constraint of grid voltage maintaining transient stability under expected severe fault. It has the objective of reducing the impact on grid active power loss caused by generator reactive power. A regulation method of using power flow calculation to optimize the steady-state reactive power output of the sensitive generator is proposed. Finally, the validity of the proposed method is verified based on the PSASP calculation model for Yahu DC transmission fed-in Jiangxi power grid.