Productivity Evaluation Method for Offshore Thick–Thin Interbedded Reservoirs Based on Graph Attention Multilayer Perceptron

Offshore multilayer sandstone reservoirs are characterized by complex vertical alternating thick and thin layers, resulting in significant heterogeneity. Traditional productivity evaluation methods often fail to effectively represent the dynamic production patterns of individual wells. This study focuses on the S oilfield offshore (Bohai Bay, China) as a case study. By considering the structural characteristics of thin layers and sand bodies, the reservoir is classified into four types: strong continuous thick layers, weak continuous thick layers, alternating thick–thin layers, and weak continuous thin layers. Based on this classification, a multilayer perceptron classification model based on graph attention neural networks is developed. The model achieves a high classification accuracy of 96.6% by mining the interdependencies between 14 input parameters. Further, by fitting the relationship between interlayer interference coefficients and water cuts for typical wells, a dynamic variation diagnosis plot for interlayer interference coefficients under different reservoir combinations is established. Additionally, a calculation method for the oil productivity index based on reservoir combination patterns is proposed. The method’s effectiveness was validated through field application, where the results significantly improved the correlation between the water-free oil productivity index and flow coefficient, with calculation errors of less than 10% compared to measured values.