Due to the lack of high temporal and spatial resolution observations, the diurnal variation of the trapped lee wave (TLW) remains unknown. We employed the U-Net deep learning model to identify more than 3.5 thousand images with TLWs from over thirty thousand 500-m resolution Fengyun-4 satellite images with a 15-minute interval during the winter times from 2019 to 2023. Results shows that the wavelength is peaked at the late afternoon with high low-level winds and low atmospheric stability while the amplitude and propagating area are peaked at relative earlier afternoon with a most turbulent boundary in a day. The TLWs were further investigated in realistic and idealized large-eddy simulations by using WRF and ICON models, respectively. It is found that the effects of the stagnant and stable layer near the surface plays a wave-absorbing role in the nonlinear regime as in linear theories or simulations while the wave lengthening is largely related to the wind speed over the mountain.