In thereconstruction of actual subsurface structures, strong noise limits the accuracy of the magnetotelluric (MT) method,causing adverse effects on later data interpretation. Given this and the characteristics of the MT time series,this study analyzed different types of noise in the MT time series,proposing a signal denoising technique based on variational mode decomposition (VMD) and long short-term memory (LSTM) predictive reconstruction. First, baseline drift correctionwas performed for the original MT datausing the VMD signal decomposition algorithm. Then, the time series was further decomposed into multiple different intrinsic mode functions (IMFs) through VMD. The LSTM time series detection model was trained using interference-free data in the RSE component, which was then identified. Afterward, the time intervals containing noise weremarked, the increasement of noise was calculated, and the noise information wastransmitted to the original signal for truncation and removal. Finally, an LSTM multi-dimensional prediction model was trained for the IMFs, followed by the prediction of missing values under various modes. The predicted results under all modes were combined to obtain the final predicted MT signals. After signal reconstruction, a secondary signal-noise separationwas performed for spike-pulse noise that was not effectively identified through VMD. TheVMD-LSTM-based signal denoisingtechnique can accurately identify strong noise in MT signals by merely processing the time series intervals containing noise, thuseffectively preserving interference-free data. Moreover, its prediction errors can berestricted within the allowable error range of the data processing for MT signals. Therefore, this technique enjoys significant denoising effects.