As a critical target in global oil and gas exploration, deep tight sandstone reservoirs present significant challenges for conventional identification methods due to their low porosity, low permeability, and strong heterogeneity. Building upon conventional convolutional neural network models, this study introduces a fully connected network architecture incorporating geological constraints such as stratigraphic grids and seismic phase information to resolve sample conflicts. A geologically oriented well screening method is proposed, which effectively accommodates diverse sedimentary microfacies and lithological combination patterns. This approach enhances sample diversity and improves reservoir parameter prediction accuracy, providing technical support for tight sandstone reservoir prediction and “sweet spot” distribution analysis.