The invention provides a frequency modulation capacity time-sharing optimization method based on conditional probability, and belongs to the field of power system automatic power generation control of. The method comprises the steps of firstly collecting the historical data of an AGC control area, and screening a sample composed of the historical data according to an AGC assessment period; constructing and training an extreme learning machine model predicted by a net load standard deviation section to obtain a trained extreme learning machine model; in an application phase, outputting, by the trained extreme learning machine model, the net load standard deviation section predicted values corresponding to the respective time periods on a certain day in the future, and according to the screened data of each AGC assessment period, calculating frequency modulation performance standard-reaching probabilities corresponding to the up-regulation capacity and the down-regulation capacity of the prediction time period, and obtaining the up-regulation reserve capacity optimization result and the down-regulation capacity reserve capacity optimization result of this time period. The method can correct the calculation result of a frequency modulation capacity demand according to a frequency modulation score, and the obtained result can truly reflect the frequency modulation capacity demand of a power system.