AVO inversion, based on the Zoeppritz equation, extracts various hidden petrophysical parameters from pre-stack seismic data. In seismic data, angle information is recorded in the form of offset values, and converting between offset values and angles is prone to generate errors. In addition, using the same approximate formula for different acreage types may lead to reduced applicability due to varying actual geological conditions. The exact Zoeppritz equation will lead to increased computational demands due to its high complexity. Therefore, this study developed an adaptive prestack inversion method based on the quadratic Encoder-Decoder network. This inversion method used the high feature and relationship extraction abilities of deep learning to replace traditional relationships, thereby reducing the angle errors and adapting to varying acreage types and geological conditions. The quadratic Encoder-Decoder network used a quadratic algorithm as the optimization method, maximizing the efficiency of the standard Encoder-Decoder structure. Additionally, the Xavier initialization method was incorporated to enhance the randomness of model initialization, thus improving the robustness of the network. The results indicate that the quadratic Encoder-Decoder network, selected through orthogonal experiments, outperforms the single-decoder network in prediction and exhibits greater consistency with actual log curves. The P-wave velocity, S-wave velocity, and density profiles obtained from inversion are consistent with the geological conditions of the study area, exhibit strong lateral continuity, and can effectively achieve high-precision prestack inversion.