基于双向长短期记忆神经网络的三维地应力场模拟

打开文本图片集

Abstract： The accurate prediction of in situ stressfields is crucial for designing hydraulic fracturing operations， as theydirectly influence fracture propagation and overallproductioneficiency.Traditional cokriging modeling methods often struggle to capture thecomplex nonlinearrelationships between multiple geomechanical parametersand seismic atributes，especially when these parameters are influenced by depositional environments or lithological variations， resulting in reduced prediction accuracy.To addressthese limitations，this study proposes a novel method for threedimensional in situ stressfield predictionusingabidirectionallong short-term memory （Bi-LSTM） neural network. Thismethod constructsathree-dimensional geomechanical constraint model，whichisintegratedwiththefinite element method for stress field prediction. Here， we preprocesswell-logging data， geomechanical parameters，and seismic atributes，and train the Bi-LSTMmodel tocapture thecomplex spatial corelationsamong these parameters. The three-dimensional body constrained by the Bi-LSTM model serves as a secondary variable in the cokriging method，which，whencombinedwiththefiniteelementframework，formsacomprehensivegeomechanicalmodelthat is used to simulate the thre-dimensional stres field.The results demonstrate significant improvements in prediction accuracyand reliability，with mean absolute error reductionsexceeding 80% and improvementsin goodness of fitby over 7% ，compared to traditional recurrent neural network methods.The mean prediction errors for the maximum horizontal stress，minimum horizontal stress，and stress orientation are 2.29% ， 2.19% ，and 7.97% ，respectively.These findings suggest that the proposed approach not only enhances the accuracy of in situ stressfield predictions but also offers a novel reference for applying machine learning methods to stress field simulation， potentially advancing the development of more effective hydraulic fracturing designs.

Key words：current stress field simulation；wellogging interpretation；bidirectional long short-term memory neural network； Cokriging method; machine learning

准确描述地应力场对于水力压裂设计决策具有重要意义，它直接影响压裂缝的扩展效果（NicksiarandMartin，2012；魏海峰，2023；汤继周等，2023；张培先，2023；陈珂等，2023）。（剩余24663字）