Using EMD method to denoise the low frequency oscillation signal of the power system with low signal-to-noise ratio, there will be a large error, affecting the identification accuracy of low-frequency oscillation signal. In order to solve these problems, an improved EMD denoising method is proposed in this paper. The   IMF modal components are obtained by EMD decomposition of the low frequency oscillation signal of the power system, the normalized autocorrelation functions are obtained and the demarcation point   between noise dominant mode and signal dominant mode is determined. Then, the noise dominant mode is denoised, and the denoised components are reconstructed with the signal dominant mode to obtain the power system low frequency oscillation signal. Finally, multi-signal Prony analysis of reconstructed signals is carried out to extract the characteristics of low-frequency oscillation in power system. The experimental results show that the improved EMD method is more effective for low frequency signal denoising with low signal-to-noise ratio (SNR), the improved EMD method and the multi-signal Prony algorithm are applied to improve the performance of the EMD multi-signal Prony algorithm. The characteristics of low-frequency oscillation signal in power system have the advantages of fast speed, high resolution and good fitting effect.