Load fluctuation is an important factor leading to frequency instability. Short-term load forecasting can predict the active power output of the system and suppress the frequency fluctuation effectively. In order to solve the problem of frequency fluctuation after asynchronous connection of Southwest Power Grid, the probability distribution of load fluctuation amplitude in Southwest China is firstly analyzed. Secondly, the load fluctuation probability is incorporated into load forecasting process to improve the forecasting accuracy. Then, based on the load probability distribution characteristics of Southwest Power Grid, a new hybrid algorithm combining back propagation (BP) neural network and support vector regression (SVR) is proposed. BP neural network is used to evaluate the load level, and the load level obtained is used as the selection basis of SVR algorithm training set to predict the load results of the day. Finally, based on the actual load data, the frequency simulation is carried out. The prediction results are compared with the traditional prediction methods without considering the load probability distribution characteristics, which verifies the superiority of the proposed model in prediction accuracy and effectiveness.