Abstract: Although the online inertia estimation method based on system identification can estimate the inertia of the power system, it is unable to determine the best model order, leading to large errors. In order to improve the accuracy of online estimation of inertia, this paper first preprocesses the measured data under the normal operating conditions of a power system to prevent over-fitting caused by noise information and to improve the ability to resist noise disturbance. Secondly, the generator’s active power change is taken as input, and frequency fluctuation is taken as output. The AIC criterion is used to determine the order of the system identification model, and the system identification method is used to establish a suitable dynamic model. Then the step response of the identification model is used to calculate the inertia of the system to avoid errors caused by the reduction of the model in the inertia online evaluation process. Finally, a simulation system is built in Matlab/Simulink to verify the effectiveness and accuracy of the proposed method.