Research On Reservoir Lithofacies Classification Based On Statistical Learning Method

The analysis of the distribution of rock facies in the formation can help us understand the geological structure of the work area.It is very important for the design of early oilfield development programs and the optimization of the development process.Reservoir facies is a comprehensive reflection of a geological region.Usually,log information can be used to obtain facies information at the location of reservoir wells.However,due to the high cost of drilling,the logging data in actual work is extremely limited,and the distribution of the reservoir's lithofacies can not be obtained through the logging data.At this time,the seismic data can be calibrated by logging data and then used.The seismic facies obtained by the seismic data indirectly reflects the plane spread of the reservoir facies.Therefore,the reservoir facies analysis is mainly divided into two parts.One is to obtain the reservoir facies directly from the well logging data and the second is to perform seismic facies analysis based on the seismic data.In the conventional conventional well logging data processing,the artificial discrimination and combined with the logging data are mainly used to classify reservoir facies.However,this method has many problems such as large manpower consumption and large man-made errors.In addition,the log data has a high dimensionality and data distribution characteristics.Complex,in most cases it is very difficult to accurately identify using some method that obeys a specific distribution.For seismic data,commonly used well logging data obtain seismic phase classification labels and then perform classification processing.Because the number of labels is small,the accuracy of seismic phase classification results is not high.To solve the above problems,this paper introduces several typical statistical learning methods to improve the accuracy of lithofacies in well logging data and solve the problem of seismic facies classification under small sample conditions.The main work of this paper is as follows:(1)According to typical lithofacies classification data of a certain working area in southwestern China,the characteristics of reservoir lithofacies data are intuitively analyzed using histograms,intersection maps,and other methods.Then,through the Kmeans clustering method and combined with the methods such as HQ-PCA,Sparse Autoencoder and other dimensionality reduction methods,the intrinsic structural features of the lithofacies classification data are analyzed,which are the reservoir rocks of the later text.Phase classification lays the foundation.At the same time,we also found that the use of conventional unsupervised clustering methods is not accurate for prediction of lithofacies and seismic facies,and the prediction results lack physical meanings.We need to further adopt supervised classification methods for the classification of well facies and seismic facies.(2)Based on the analysis of the previous logging data,we introduced ADABOOST ensemble learning method and adopted SVM and ANN as basic classifiers for supervised logging and lithofacies classification.Through quantitative analysis and comparison of the traditional SVM and ANN classification results,it is found that the proposed method has higher classification performance than traditional methods,can handle data that cannot be correctly classified by conventional methods,and has stronger stability and generalization ability.(3)For seismic data,the seismic phase classification has fewer label samples.This paper introduces a semi-supervised process flow based on conditional random fields for seismic phase classification.The training samples are extended by conditional random fields and the maximum correlation entropy criterion,and a multi-person decision making method is used to obtain a seismic phase distribution with probabilistic significance.Comparing synthetic seismic data classification results,it is found that the method proposed in this paper can obtain more accurate seismic facies classification results than the above-mentioned supervised methods such as SVM when the sample data is 1%.Finally,this method is applied to actual data and cross-validation proves the effectiveness of the proposed method.In summary,this paper has studied the two key contents of the logging and lithofacies classification and seismic phase classification in the reservoir lithofacies classification.After deeply analyzing the distribution characteristics of logging data and seismic data,an integrated learning method based on ADABOOST framework was proposed to solve the problem of multi-distribution coupling of well-facies lithofacies classification data;Semi-supervised seismic facies classification method based on conditional random fields and maximum correlation entropy criterion,and the seismic facies classification results with probabilistic significance.The research contents of this paper can provide ideas and lessons for the application of machine learning methods in the field of geophysics,and have a positive effect on promoting the intelligent development of oil and gas exploration technology.