Physics-constrained distributed neural network model for 3D in-situ stress prediction
In-situ stress is an important basic parameter in the whole life cycle of an oil and gas reservoir,such as wellbore stability analysis,hydraulic fracturing design,sand production prevention,casing damage prediction and prevention,and oil and gas exploitation measures.Aiming to address the challenge of predicting 3D in-situ stress fields in the absence of 3D seismic data,a physics-constrained distributed neural network model(PDNN)was proposed for 3D in-situ stress prediction.This model was developed based on the logging interpretation results of in-situ stress for drilled wells in the working area.Firstly,the single well profile of in-situ stress was obtained by logging interpretation method,and the 3D geological model was constructed by Kriging interpolation.Secondly,the logging data and the 3D spatial coordinates of the 3D geological model were input into three fully connected neural networks.The physical constraints of the in-situ stress were introduced,the conditions of both the data and the physical constraints were proposed,and the 3D spatial coordinates were employed to forecast the 3D in-situ stress field.Subsequently,the weight parameters of both the data and the physical constraints were selected and compared with three machine learning models,including the artificial neural network(ANN),support vector regression(SVR),and random forest(RF),as well as the Kriging interpolation model.Finally,the effect of different machine learning models in predicting in-situ stress was then evaluated.The results indicated that:(1)The in-situ stress in the study area was consistent with the potential normal faulting stress state,namely,vertical principal stress＞maximum horizontal principal stress＞minimum horizontal principal stress.(2)The value of the weight parameter,which incorporates both the data and the physical constraints,had a significant influence on the prediction results.The optimal prediction performance for the 3D in-situ stress was achieved when the weight parameter was set to A=0.2.(3)In comparison to the ANN,SVR,RF,and Kriging models,the 3D in-situ stress and pore pressure predicted by this model exhibited greater accuracy.The average absolute percentage errors of the vertical principal stress,maximum horizontal principal stress,minimum horizontal principal stress,and formation pore pressure predicted by the testing set in the working area were 0.63％,7.59％,7.16％,and 3.21％,respectively.Furthermore,the model proposed in this paper was capable of accurately capturing the 3D variation characteristics of in-situ stress gradient.The normal distribution characteristics of in-situ stress and pore pressure in horizontal wells were readily apparent.It can be concluded that the PDNN model is an effective means of integrating the relationship between topography and in-situ stress,thereby improving the accuracy and interpretability of 3D in-situ stress prediction,and it provides a novel approach to the prediction of 3D in-situ stress fields in oilfields.
3D in-situ stressPore pressureNeural networkPhysical constraintData-drivenGeological model
万方数据
维普
