Intelligent Inversion Of Petro-physical Parameters In Presence Of Few Labels

Petro-physical parameters refer to parameters such as porosity and permeability that describe the physical properties of the reservoir.These parameters can accurately and quantitatively describe the pore size of the reservoir and the ability of the reservoir to allow fluid to pass.Thus,it has important applications in oil and gas development.The main method of obtaining accurate reservoir physical parameters is to take core samples from wells and then measure them through experiments.This method is too expensive and cannot be performed in locations where there is no logging and cannot obtain the prediction results of physical properties of the entire work area,so its use is limited.There are two main methods for predicting traditional reservoir physical parameters: 1.Establish a linear regression relationship between reservoir elastic parameters and physical parameters through petrophysical models to predict physical parameters;2.Establish deep learning models to predict reservoir physical parameters.The former lacks consideration of reservoir pore structure and lithofacies,which leads to deviations in predicting physical property parameters,while the latter requires a large number of samples during the training process due to the complexity of the model,and sample acquisition is costly and difficult to satisfy.Therefore,the model has small overfitting and other problems.Potential problems caused by the sample.To solve the above problems,based on the facies control method and active learning algorithm framework,this paper constructs a reservoir physical parameter prediction model under the condition of small samples.The main research contents are as follows:(1)Introduced pore structure parameters and proposed the phase-controlled regression method.Traditional petrophysical models only consider the effect of elastic parameters on physical parameters when predicting reservoir physical parameters,while ignoring the influence of pore structure and rock control on physical parameters.This paper introduces the framework flexibility factor to quantitatively characterize the pore structure of the reservoir.At the same time,this paper constructs a phase-controlled physical parameter prediction method.Experiments show that the framework flexibility factor can effectively improve the prediction accuracy of reservoir physical parameters.The facies-controlled reservoir physical-parameter prediction method uses lithofacies as a constraint to establish multiple physical-parameter regression models,and fully considers the impact of lithofacies in the physical-parameter prediction.Thus,the prediction error of reservoir physical parameters is reduced.(2)Constructed a lithofacies classification model with few samples.Considering the particularity of logging data and seismic data,we conducted detailed experiments on the selection of query strategies and classification models in the active learning framework.The results show that the selection of query strategies in different case is very important.The optimized query strategy has better performance in complex multi-classification scenarios.The selection of the basic classification model depends on the specific conditions of the data.The more complex the model,the better the performance,but the more queries are required.Active learning framework can reduce the number of samples,effectively solves the few sample problem.(3)Constructed a regression method of petro-physical parameters under the condition of few samples.Based on the existing active learning regression query strategy,phased regression method is improved based on the representativeness and diversity of the sample.Through a small number of training samples,a high-precision phasecontrolled regression model can be constructed.Experiments show that there are large errors in the prediction results of petro-physical parameters with few samples,and the amplitude of the inversion results is unstable.In the active learning framework,only a small amount of learning is needed to greatly improve the accuracy of the phasecontrolled physical-parameter inversion model.