The accuracy of low-frequency model in seismic inversion plays an important role in improving the quality of inversion results, especially for lacustrine sedimentary strata during the Paleogene with complex tectonic faults or fast variation of sedimentary facies. The development of conventional low frequency model will produce “bull's eye” phenomenon due to well interpolation, or increase the uncertainty of inversion results due to unreasonable seismic velocity. In order to solve the above problems, taking LA oilfield with complex structural and sedimentary background as an example, band-limited P-wave impedance, P/S velocity ratio and density data volume can be obtained by optimizing the issues of residual moveout and near-offset multiples in the original gathers and carrying out pre-stack simultaneous inversion based on constant model; at the same time, due to the error between seismic velocity and the low frequency trend of the well velocity, seismic velocity must be corrected; and then, the band-limited inversion data volume, corrected seismic velocity volume and logging data are incorporated using the theoretical method of deep learning to obtain the full-band P-wave impedance, P/S velocity ratio and density inversion data volume, which are used as the low-frequency model for subsequent inversion. The study shows that the low-frequency model based on deep learning can well reflect the variation characteristics of sedimentary facies, and improve the accuracy and predictability of inversion results; it can also provide better characterization of hydrocarbon distribution in reservoirs and define the distribution of major payzone in the reservoir, providing strong technical support for the implementation of subsequent development projects in LA oilfield of the Pearl River Mouth Basin.