Research On Reservoir Prediction Method And Application Based On Deep Learning And Seismic Data

Natural gas is a strategical-clean-energy with short supplement in China.To increase the exploration of natural gas is an important part of the national energy strategy.The traditional reservoir prediction method has good application effect in shallow exploration,but it shows inapplicability in deep exploration.This paper aims at the deep exploration targets,introduces the theory and methods of “deep learning” in the research field of artificial intelligence,and extracts the gas-responsive characteristics of seismic data by constructing a deep learning unsupervised feature learning network model.The carbonate natural gas reservoir of the Proterozoic Sinian Dengying Formation in Central Sichuan Province have the characteristics of deep burial,long evolutionary time,and complex reservoir characteristics.This paper uses the region as an application object to carry out the research on reservoir prediction methods and application based on deep learning and seismic data.The main research contents and achievements of this article are as follows:1.Constructing a deep learning network model suitable for seismic data is conducive to extracting gas-responsive features.Seismic data belongs to real-valued data and has continuity in time and space.In this paper,a continuous restricted Boltzmann machine(CRBM)with simulated continuous data probability distribution is introduced as a shallow model of deep network in combination with the characteristics of seismic data.The deep network has the capability of layer-by-layer abstraction.This paper builds a deep belief nets(DBN)by stacking CRBMs and symmetrically expands to obtain a deep autoencoder.Before training,a proper amount of Gaussian noise is added and the original data is used for reverse fine-tuning to make the network output get as close to the original data as possible.Through the above method,a Deep-CRBM-Denoising AutoEncoder(DCDAE)using CRBM as shallow model was finally constructed.2.The optimization of the target features of the deep unsupervised feature learning network needs to be evaluated in conjunction with visualization techniques and feature analysis methods.High-level features of deep learning are highly abstract.In order to analyze the connection between high-leve features and goals,high-level features need to be visualized for comparative analysis.This paper introduces deep learning feature visualization technology and applies it to high-level feature visualization of seismic data.After the deep network completes the unsupervised feature learning,it often obtains a large number of features.In order to screen out the target response features,a SOM topology feature analysis method is proposed.According to the category distribution of the target sample and the visual features of the same hidden layer on the topological map,the favorable features are filtered out to facilitate the further fine description of the gas distribution.In order to verify the feasibility of a series of methods,the Chuanzhong forward model was constructed.The network training and the optimization of the target characteristics were completed in combination with relevant methods.The specific impact of the target characteristics and other category characteristics on the research data was compared and analyzed.3.For the study area,one-dimensional seismic data and three-dimensional seismic data were used to conduct deep learning gas detection methods.Cluster analysis is an important method to analyze the relationship between different data objects.Kmeans algorithm is simple,practical and easy to implement.It is limited to the random initialization of the center point,and the clustering effect is not stable.This paper uses SOM's unsupervised clustering method to obtain relatively stable clustering results.After completing the deep network training,compare the category distribution of the original data clustering and all the feature activation clustering of different hidden layers,and infer the hidden layer level of the target features.According to the SOM topology feature analysis method,the target features of the corresponding hidden layer are optimized,and the clustering method is used to determine the optimal target features and divide the gas-bearing interval of the study area.The results show that both training modes can achieve the effect of gas detection.In comparison,the gas-bearing test results of three-dimensional seismic data show better continuity and more clearly divide the gas distribution of the study area.