The large fluctuation of wind power and the poor accuracy of day-ahead prediction lead to the contradiction between large-scale wind power accommodation and grid security robustness. An automatic frequency modulation scheduling method for wind power "waist load" output access is proposed to reduce wind curtailment and improve the online security of the main network under large fluctuations of wind power. Firstly, to improve the convergence of neural network for day-ahead wind power prediction, a multi-sample data preprocessing method based on singular value decomposition of time series matrix is proposed. Secondly, in order to obtain the output plan curve of waist load related to the day-ahead wind power prediction, polynomial regression fitting and reference power deviation are used to obtain a smooth curve with high correlation. Then, to reduce the frequency deviation caused by the sudden decrease of online wind power, a family of "start-up" curves for automatic frequency modulation of frequency modulation units is set. Finally, in order to maintain a reasonable active power flow distribution in the main network after automatic frequency modulation and improve the power angle security, an optimization model with the smallest equivalent power angle is adopted to obtain the optimal allocation scheme of the output increment of each frequency modulation unit. The example verifies the feasibility of the scheduling method, which has theoretical and practical significance for reducing wind curtailment of large-scale wind power and improving the safe operation level of the power grid.