Accurate construction of a seepage model for a multifractured horizontal well in a shale gas reservoir is essential to realizing the forecast of gas well production, the pressure transient analysis, and the inversion of the postfracturing parameters. This study introduces a method for determining the fracture control region to characterize the flow area of the matrix within the hydraulic fracture network, distinguishing the differences in the flow range of the matrix system between the internal and external regions caused by the hydraulic fracture network structure. The corresponding derivation and solution methods of the semi-analytical seepage model for fractured shale gas well are provided, followed by the application of case studies, model validation, and sensitivity analysis of parameters. The results indicate that the proposed model yields computational results that closely align with numerical simulations. It is observed that disregarding the differentiation of matrix flow area between the internal and external regions of the fracture network led to an overestimation of the estimated ultimate recovery, and the boundary-controlled flow period in typical well testing curves will appear earlier. Because hydraulic fracture conductivity can be influenced by multiple factors simultaneously, conducting a sensitivity analysis using combined parameters could lead to inaccurate results in the inversion of fracture parameters.

Accurate estimated ultimate recovery prediction of fractured horizontal wells in tight reservoirs is crucial to economic evaluation and oil field development plan formulation. Advances in artificial intelligence and big data have provided a new tool for rapid production prediction of unconventional reservoirs. In this study, the estimated ultimate recovery prediction model based on deep neural networks was established using the data of 58 horizontal wells in Mahu tight oil reservoirs. First, the estimated ultimate recovery of oil wells was calculated based on the stretched exponential production decline model and a five-region flow model. Then, the calculated estimated ultimate recovery, geological attributes, engineering parameters, and production data of each well were used to build a machine learning database. Before the model training, the number of input parameters was reduced from 14 to 9 by feature selection. The prediction accuracy of the model was improved by data normalization, the early stopping technique, and 10-fold cross validation. The optimal activation function, hidden layers, number of neurons in each layer, and learning rate of the deep neural network model were obtained through hyperparameter optimization. The average determination coefficient on the testing set was 0.73. The results indicate that compared with the traditional estimated ultimate recovery prediction methods, the established deep neural network model has the strengths of a simple procedure and low time consumption, and the deep neural network model can be easily updated to improve prediction accuracy when new well information is obtained.