We apply the variational autoencoder (VAE) to the LAMOST-K2 low-resolution spectra to detect the magnetic activity of the stars in the K2 field. After the training on the spectra of the selected inactive stars, the VAE model can efficiently generate the synthetic reference templates needed by the spectral subtraction procedure, without knowing any stellar parameters. Then, we detect the peculiar spectral features, such as chromospheric emissions, strong nebular emissions, and lithium absorptions, in our sample. We measure the emissions of the chromospheric activity indicators, H alpha and Ca ii infrared triplet (IRT) lines, to quantify the stellar magnetic activity. The excess emissions of H alpha and Ca ii IRT lines of the active stars are correlated well to the rotational periods and the amplitudes of light curves derived from the K2 photometry. We degrade the LAMOST spectra to simulate the slitless spectra of the China Space Station Telescope (CSST) and apply the VAE to the simulated data. For cool active stars, we reveal a good agreement between the equivalent widths of H alpha line derived from the spectra with two resolutions. The result indicates the ability of identifying the magnetically active stars in the future CSST survey, which will deliver an unprecedented large data base of low-resolution spectra as well as simultaneous multiband photometry of stars.