Determine relative permeability curves for gas reservoirs based on machine learning
Relative permeability curve is fundamental for investigating multiphase seepage,and it is of great significance in calculat-ing the gas well production and analyzing the water-producing law.Conventional relative permeability curves are generally obtained through core experiments,which are time-consuming and costly.Therefore,this paper proposes a method for calculating gas-water relative permeability curves using production performance data,based on integrating empirical formula,reservoir numerical simu-lation,and machine learning.Taking a study block of a gas reservoir in the Sichuan Basin as an example,a sample set consisting of gas production,water production and formation pressure calculated by numerical simulation was utilized as the model input and the parameters of the Brooks-Corey model as the output.By comparing the learning algorithms of the Scaled Conjugate Gradient(SCG),Bayesian Regularization(BR),and Levenberg-Marquardt(LM)neural networks,a training model was established by the selected LM algorithm.Furthermore,the influence of the size of the sample set on prediction results was discussed and the corresponding optimiza-tion strategy proposed.The following results are obtained.(i)Different hidden layer settings can cause different training effects on the neural network,the best prediction of relative permeability curve is achieved when the LM algorithm is adopted with 2 hidden layers which have 41 and 32 nodes respectively.(ii)The sample size has great influences on the network training speed and model prediction effect.Appropriate reduction of the samples can improve the model prediction effect,but will increase the network training error.Re-duction in the input variable time period will increase the network training error and reduce the final prediction effect of model.Prac-tical application demonstrates that the difference between the predicted relative permeability curves and the core relative permeability curves is small,while the fitting accuracy of water production and pressure is high.This method is proved to be efficient and accurate for calculating gas-water relative permeability curves,providing a strong support and guidance for the development of gas fields.
Relative permeability curveMachine learningArtificial neural networkNumerical simulationBrooks-Corey model
国家科技期刊平台
NSTL
万方数据
