Since traditional Prony analysis of low frequency oscillations has strict requirements to the input signal and is sensitive to the noise of data, this paper proposes a empirical mode decomposition filtering and Prony analysis combined method for low frequency analysis. In this method, empirical mode decomposition is used to adaptively filter the noise of the input signals before improved Prony analysis is carried out. The order of improved Prony analysis is determined by the normalized singular value method. This method is applied to analyze the test signal and the IEEE 4-machine system oscillation signals, and compared with the Prony analysis based on low pass filter. The simulation shows the effectiveness of this method which indicates that the result of the analysis is good even in highly noisy environment.