Abstract: The secondary frequency regulation or automatic generation control of a power system requires a certain capacity of frequency regulation reserve. This paper proposes a data-driven based method for estimating frequency regulation reserve capacity (FRRC) requirement using conditional probability. First, the correlations between the control performance and its affecting factors are analyzed and established in the form of conditional probability. Secondly, in order to predict the power fluctuation, an interval model for predicting the net load standard deviation is established based on the extreme learning machine. Finally, a data-driven approach is proposed to optimize the FRRC based on the established relationship and the predicted net load fluctuation. The numerical simulation results based on the historical data show that the proposed method can reduce FRRC and improve the control performance.