Abstract: The semi-analytical model of galaxy formation in L-Galaxies has successfully predicted many galaxy properties. In this paper, we use L-Galaxies to simulate physical results such as the star formation history of galaxies. Considering the dust attenuation and re-radiation features, combined with the results SKIRT (Stellar Kinematics Including Radiative Transfer) radiative transfer simulation as the training data of the deep learning model starduster, we can predict galaxies' multi-band spectral energy distribution. The dust extinction of starduster is slightly insufficient, resulting in mildly higher than the observation data from FUV (far-ultraviolet) to optical, and at the longer wavelength, the number density of galaxies at the bright end is well reproduced, but under-predict the number of faint galaxies slightly. However, taking the empirical dust extinction combined with the infrared spectral energy distribution template as the empirical model to obtain the luminosity function from the FUV to the FIR (far infrared), the dust attenuation effect is quite good. And we find it in excellent agreement with observation at the UV (ultraviolet) band. Except that the number of galaxies at the faint end cannot be reproduced well, the luminosity functions at infrared are predicted accurately.