Effective use of the difference between longitudinal and converted shear waves in reservoir sensitivity analysis is a key component to reduce the multi-solution problem of using longitudinal waves only in reservoir characterization. Deep neural network (DNN) is proven to be a powerful tool to solve an end-to-end task in a purely data-driven way, which brings a promising potential in reservoir characterization applications. In this paper, we design a reservoir prediction method using cluster analysis and DNN method. Seismic attributes sensitive to oil and gas responses are first optimized by cluster analysis, and then, a multi-component composite operation is carried out on the optimized attributes to extract oil and gas characteristics. Finally, the DNN is tested and trained to determine the best network model. The final DNN model is further examined using multi-component data in the Fenggu structural area (Sichuan, China) for seismic gas reservoir prediction. The results show that the seismic gas reservoir distribution predicted using this scheme is generally consistent with actual drilling information. Compared to single-component data, the multi-component composite seismic attributes trained network provides prediction results with higher accuracy and reduces the uncertainty of inversion results.