LAUSR

Abstract Tight oil and gas reservoirs with huge development potential are widely distributed all over the world and horizontal well technique is a popular development technology for such kinds of reservoirs. However, the efficiency of conventional horizontal well technology is usually unsatisfying due to the unfavorable flow conditions caused by the nature of tight reservoirs, such as low matrix permeability and narrow pore throat. Multi-stage fractured horizontal well technology is an emerging attractive technology that can greatly improve oil production by generating highly conductive fracture networks in tight reservoirs. and numerical simulation method is generally used to predict the production dynamics of the multi-stage fractured horizontal wells. Nevertheless, the complex geological conditions of reservoirs would greatly increase the difficulties and time required to implement a complete simulation (including model building and calculation). Therefore, in this paper, deep convolution generative adversarial networks (DCGAN) based on U-Net framework was applied to establish the dynamic mapping relationship between fracture pattern and reservoir pressure during the production process, that is, the dynamic reservoir pressure distribution can be obtained by image mapping the fracture distribution details. The results showed that the efficiency of U-Net framework based deep convolution in extracting, dividing and splicing the geometric features of fracture network is significant, and the generative adversarial networks model can effectively predict the reservoir pressure distribution according to the fracture network geometry after being trained with 6000 sets of data. Herein, the training accuracy of the proxy model towards sample data was compared, and the confrontation between the generator and the discriminator in the iterative training process was clarified according to the error function. Furthermore, the prediction accuracy towards pressure distribution at different fracture network geometry scales by the proxy model was compared. Results indicated that the pressure diffusion range predicted by the proxy model is basically consistent with the range obtained from the numerical simulation, with a mean square error (MSE) generally smaller than 0.2. In addition, the relationship between the prediction accuracy of the proxy model and the number of sample data and iterative training times was also studied, which showed that the mean square error of the proxy model kept decreasing with the increasing iteration cycles and sample size, which met the basic law of statistical learning. Moreover, it is worth mentioning that the proposed method may also shed light on the applications of DCGAN in other reservoir-related problems.