Proxy Model Forecasts Production of Shale Wells With Complex Fracture Networks

Abstract

Listen

Translate article

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 209635, “A Deep-Learning-Based Approach for Production Forecasting and Reservoir Evaluation for Shale Gas Wells With Complex Fracture Networks,” by Peng Dong, SPE, and Xinwei Liao, China University of Petroleum at Beijing. The paper has not been peer reviewed. _ The complete paper proposes a data-driven proxy model to effectively forecast the production of horizontal wells with complex fracture networks in shales. With a multilayer gated recurrent unit (GRU) cell, the proxy model is coupled with newly developed deep-learning methods, an attention mechanism (Att-GRU), skip connection, and cross-validation to deal with time-series-analysis (TSA) issues of multivariate operating and physical parameters. Results indicate that the Att-GRU method can forecast the production for shale gas wells with complex fracture networks accurately at a given time and with variable bottomhole pressure (BHP) while maintaining high calculation efficiency. Methodology The authors write that their main interest is to develop a physics-based data-driven model to speed up the simulation of complex fracture networks in shale reservoirs. Results show that the deep-learning model can faithfully recapitulate the variable BHP condition in all of its intricacies with multivariate physical input. In this section of the complete paper, the involved physical background and the corresponding governing equations are clarified. Then, the principle behind the Att-GRU for the problem of complex-fracture-network wells in shale reservoirs is explained. Data initialization and preprocessing and model structure are discussed. Because these sections are equation-intensive and are not suitable for this synopsis but are essential to the methodology explained in the paper, the reader is encouraged to review them in the complete text. Results and Discussions In this section of the complete paper, the physical data in the boundary element method (BEM) are used to train and test the Att-GRU model. The training procedure, test results, and parameters analysis are clarified. The computation burden of the data-driven Att-GRU is compared with the physical model.