Modern power system has become complex and the size of the nonlinear systems makes the traditional analysis method present more and more limitations. Therefore this paper introduces an improved multi-signal matrix pencil algorithm suitable for low-frequency oscillations modal identification of power system. First, the method of singular value decomposition (SVD) is used to separate the signal from the noise sub-space, confirm the order and eliminate the signal noise. And then through the establishment of multi-signal normalized experiment matrix, the matrix pencil algorithm is improved to identify power system mode. The ideal signal and simulation system are analyzed by using the original Prony method, the harmonic retrieval Prony method and the matrix pencil method. Results show that multi-signal matrix pencil method has high precision and a certain anti-noise ability. Through the normalized processing of the multi-signal, the submerging of small signals is avoided when different types of signals are overlying, thus it is suitable for low frequency oscillation on-line identification.