Denoising is an important part of magnetotelluric data processing. To enrich and develop the denoising method of magnetotelluric time series, this study introduced the LSTM network-one of the recurrent neural networks-into the square wave noise processing of the magnetotelluric time series. Different from previous studies, the measured magnetotelluric time series without human interference superimposed on simulated square wave noise were used as the input of the LSTM network, and the noise-free original time series were used as the target output of the network. After training for 1,500 epochs, the normalized cross-correlation coefficient between the time series extracted from the simulated noise signals by the network and the original time series reached 0.9718, indicating that the network has effectively learned the characteristics of the noise-free magnetotelluric time series. Finally, the denoising test results of measured square wave noise signals show that the proposed method can effectively suppress the interference of square wave noise and improve the estimation quality of impedance. This study provides a new idea for the processing of magnetotelluric time series based on deep learning.