Residual Conditional GAN for Structured Control-Pore-Volume Inversion from Dynamic Production Responses

Rapid inversion of coarse-scale parameters is important for history matching in reduced-order reservoir models, yet conventional methods require repeated simulation and iterative updating. This study proposes a residual conditional generative adversarial network (Res-cGAN) to reconstruct structured control-pore-volume parameters from well liquid production rate (WLPR), well oil production rate (WOPR), and pressure responses. The methodological contribution is a response-conditioned inversion framework that directly maps multivariate production responses to the structured CPV representation, combining residual feature learning, cross-level feature fusion, conditional adversarial learning, and local matrix-continuity regularization. A dataset of 3000 samples was generated using the interwell numerical simulation model (INSIM), with training sets of 600, 1200, 1800, and 2400 samples and fixed validation and test sets of 300 samples each. Increasing the training size reduced mean squared error (MSE) from 0.0128 to 0.0059 and increased the coefficient of determination (R2) from 0.9120 to 0.9670. With 2400 training samples, Res-cGAN achieved an MSE of 0.0059, mean absolute error (MAE) of 0.0493, mean absolute percentage error (MAPE) of 5.18%, and R2 of 0.9670, outperforming the baseline conditional GAN (Base-cGAN). These results suggest the potential of Res-cGAN for rapid post-training CPV initialization and candidate screening for reduced-order history matching workflows, while field validation and cross-reservoir testing remain necessary.