Research On Gas Reservoir Identification Method Based On Deep Learning

The gas-bearing evaluation of reservoir is the core of natural gas exploration.Most reservoir is buried in the deep ground at present,with the characteristics including weak seismic and pore fluid response,little the difference between reservoir and non-reservoir,leading to the gas detection of gas reservoir becoming difficult.After decades of constant efforts,the seismic prospecting specialists have developed some remarkable methods for hydrocarbon detection,such as blight spot technique and AVO analytical approach and so on.These techniques have both success and failure.The gas-containing evaluation of reservoir by using seismic data is still a worldwide problem.For more effectively conducting hydrocarbon detection of deeply buried reservoirs and further improving the utilization rate of seismic data to fully exploit the effective porous fluid information and further enrich the types of seismic attributes,this study introduces the automatic features extraction technology of deep learning,which is a state-of-the-art and hot field of Artificial Intelligence,into the field of seismic exploration.The goal is to mine the intrinsic seismic characteristics of the weak natural gas reservoir from the seismic data by the use of deep learning method and to identify the natural gas reservoir.The main contents and achievements of this paper are as follows:1.Researched on the deep learning model that is suitable for the seismic feature extraction of reservoirs.The Deep Belief Network(DBN),which was proposed by Hinton in 2006,is formed by the stacking Restricted Boltzmann Machine(RBM),one of unsupervised shallow learning model.RBM is a probability generation model,showing its superiority in modeling the binary image data and extracting image features,though imperfect in simulation of continuous data.Therefore,this study uses the Continuous Restricted Boltzmann Machine(CRBM)to construct a Continuous Deep Belief Network(CDBN)model,a continuous version of RBM,which can simulate the continuous data while maintain the all properties of RBM.This paper also adds sparse constraints to CRBM to generate a Continuous Sparse Restricted Boltzmann Machine(CSRBM)and build a Continuous Sparse Deep Belief Network(CSDBN).On this basis,CDBN or CSDBN is transfered to the Stacked Autoencoders to form a Continuous Deep Autoencoders neural network(CDAE)or Continuous Sparse Denoising Deep Autoencoders neural network(CSDDAE).These networks carry the whole module of deep learning feature extraction and the program.2.Proposed a selecting method of the specifying features of the deep learning categories.We can obtain multiple features after deep learning model training completed,but not all features are indicator for the target tasks(su ch as gas-bearing identification).The model will select discriminatory features by similar parameters(e.g.,correlation coefficient,distance)in the case where the obtained features are more.3.Combined the CRBM with Support Vector Machine(SVM)to form CRBMSVM model to identify the western Sichuan marine limestone and dolomite,and the correct rate reached 81.9%.Also,based on SVM,the correction rate of features recognition is higher than Principal Components Analysis(PCA)extracted by CRBM.And the features extracted by CRBM,an unsupervised shallow learning model,are more discriminative than those obtained by PCA.4.Carried out the applied features extraction based on seismic data by deep learning in the western Sichuan marine carbonate reservoir.The target zone of the study area is located at the top of Leikoupo Formation,which the seismic response is weak.In this area,there are only three wells,few logging data,difficult reservoir prediction and gas-bearing detection.Thus,the feature extraction method of three-dimensional seismic data and one-dimensional seismic data by deep learning are proposed in this paper.The clustering analysis combined with the logging data shows that the classification results accurately divide the gas and water wells.At the same time,by comparing the clustering analysis results of deep learning features from threedimensional seismic data with the results from one-dimensional data,it is considered that the class boundaries of deep learning features from three-dimensional seismic data are clearer,less noise and stronger continuity.5.Compared between high-level features clustering analysis,the shallow features clustering analysis and the original data clustering analysis processed by deep learning method.It is considered that the effect of high-level feature clustering is better than the clustering results of shallow features and the original data,and the boundary of classification is clearer.6.Selecting the deep learning top discriminatory features by clustering.The distribution of specifying features of Xiao S1,CK1 and Xin S1 matched mainly with log data.It is proved that the feature extraction method based on deep learning is feasible in this field.