It is especially important for improving power quality to identify all kinds of harmonics and inter-harmonics accurately and effectively. Due to the spectrum leakage, the traditional Fast Fourier Transformation algorithm (FFT) cannot accurately detect the non-integral harmonics. To address this issue, extended Prony algorithm is applied in this paper to identify the non-integral harmonics. Wavelet-denoising method is firstly used to overcome the weakness that the Prony algorithm is noise-sensitive, then model’s eigenmatrix is obtained by linear prediction and singular value decomposition is used to solve the eigenmatrix to get model’s characteristic polynomial, finally least square method is used to find the root of the characteristic polynomial to get parameters of the non-integral harmonics. According to the problem that the signal band width may affect the accuracy of the algorithm, extended Prony algorithm is used to analyze different bandwidth signal to prove the algorithm’s accuracy. Experiment analysis verifies the viability and effectiveness of the algorithm.