In order to solve the problems of white noise and periodical narrowband interference in PD detection, a denoising method combining improved variational mode decomposition (VMD) and threshold denoising is proposed. Aiming at the problem that the VMD algorithm is difficult to choose the decomposition parameters adaptively in the practical application, the decomposition number is determined by the principle of minimum energy deviation, the penalty factor of each component was optimized by BAS, and the kurtosis criterion was used to screen the effective component, so as to eliminate the narrowband interference noise. Threshold function combined with 3σ criterion was used to remove the residual white noise in the effective component and reconstruct the effective component. Through simulation and measured signal denoising analysis and compared with lifting db4 wavelet method and EEMD threshold method, the results show that this method has better denoising effect, the denoising waveform similarity coefficient is higher after denoising, the noise rejection ratio is higher and can retain more partial discharge characteristics.