Prony algorithm is an effective method to derive characteristics of power system low frequency oscillations. However， the traditional Prony algorithms are very sensitive to noises and can only be applied to part of signal every time. This paper proposes the moving-window Prony algorithm to the identification of low frequency oscillation of power system. With the moving-window technique， the objective function of the proposed Prony method is changed into the least square error by windows， which is different from the partial least square error corresponding to the traditional algorithms. Thus， the moving window Prony algorithm has stronger adaptability to noise and could be applied to long time range signals. The simulation results based on DSA Tools demonstrate that the moving window Prony algorithm has good and effective performance for identification of power system low frequency oscillation， even in cases when the PMU's data contains noises.