Dynamic Prediction Of Oil And Gas Reservoir Development By Deep Learning With Seepage Physical Constraints

Oil and gas reservoir development performance prediction is a process of processing and analyzing oil and gas reservoirs development performance data,re-understanding current development status of oil and gas reservoirs and predicting future development performance.Deep learning method processes historical production dynamic data by means of data approximation,which avoids the establishment of geological models and complex mathematical solution,has been widely used in the dynamic prediction of oil and gas reservoir development.However,during oil and gas reservoirs development,the cost of data acquisition is high and it is difficult to measure.At present,most deep learning algorithms do not take into account the knowledge of related fields such as oil and gas reservoir seepage physical laws,which seriously affects the accuracy of deep learning algorithm prediction.Therefore,this paper proposes seepage physical constraints deep learning algorithm,which mainly adds seepage physical laws such as oil and gas reservoirs seepage control equations and analytical solutions to the loss function of deep neural network in a regularized way.The network is physically interpretable and improves the prediction accuracy of deep learning algorithms under the condition of limited data acquisition conditions.Through research,the following understandings and results have been obtained:(1)The single-phase reservoir seepage control equation is integrated into the artificial neural network,and the artificial neural network formation pressure prediction model with reservoir seepage physical constraints is established.On the test set,the root mean square error and determination coefficient of traditional data-driven artificial neural network is 5.521 and the of is 0.804,while the root mean square error and the determination coefficient of the artificial neural network with reservoir seepage constraints are 3.815 and 0.907,so the results predicted by the artificial neural network with reservoir seepage constraints on the test set closer to the results of numerical simulations;(2)The one-dimensional water-flooding oil saturation movement equation is integrated into the artificial neural network,and the artificial neural network saturation prediction model constrained by the Buckley-Leverett equation is established.The root mean square error of the water saturation predicted by this model is 0.012,the coefficient of determination is 0.990,which verifies the validity of the artificial neural network constrained by the Buckley-Leverett equation;(3)The gas reservoir seepage control equation is integrated into convolutional neural network in the form of finite difference,and the convolutional neural network pressure prediction model with gas reservoir seepage constraint is established.The relative errors of formation pressure predicted by this model are all less than 5%,which verifies the validity of the convolutional neural network with gas reservoir seepage constraints.;(4)The basic pressure solution under the condition of constant production of natural gas reservoir is integrated into fully convolutional neural network,and a fully convolutional neural network pressure prediction model with analytical solution constraints is established.The relative errors of formation pressure predicted by the model are all less than 0.1%,the root mean square error of the model is 0.052,and the coefficient of determination is 0.997,which verifies the effectiveness of the fully convolutional neural network with analytical solution constraints.