Abstract: Partial discharge （PD） detection in power transformers is crucial for assessing insulation conditions and preventing failures. However, on-site PD signals are often contaminated by background noise, necessitating denoising to extract clean signals. To address this, an improved variational mode decomposition （VMD） dual-tree complex wavelet transform （DTCWT） denoising method is proposed. The method first employs modified VMD with a kurtosis-based criterion to preliminarily remove narrowband interference from background noise. Next, the puma optimizer （PO） algorithm optimizes the block threshold parameters for DTCWT based threshold denoising, effectively suppressing white noise. Finally, an adaptive filter eliminates residual narrowband interference. Experimental results demonstrate that, compared to conventional denoising methods, the proposed approach could effectively suppress noise, improving the waveform similarity coefficient by approximately 3% while better preserving PD signal characteristics.