An Improved SVD de-noising method combined with EMD is proposed to deal with the bad phenomena of wavelet transform in transient power quality de-noising, wavelet transform can filter out the high frequency useful signal by mistake, and resulting in serious signal distortion or excessive burr after noise elimination. In this method, the noise signal is firstly filtered by EMD, and then by constructing a Hankel matrix, using left and right singular vectors to determine the singular matrix, and then uses an improved singular value difference method to determine the rank of the effective singular matrix according to the rules of singular value distribution, thus, the signal reconstruction after de-noising is completed. The results show that this method is effective in de-noising, in particular, the effect of voltage sag and voltage interruption is the most obvious. Its signal-to-noise ratio is more than 2.9db and 7.3db respectively, the validity of this algorithm is verified.