Seismic Facies Controlled Reservoir Parameters Prediction Based On Machine Learning

With further exploration and development of oil & gas,the exploration target has changed from structural reservoir to lithologic reservoir,and the characterization of thin layer and complex layer reservoir has gradually become a difficulty in reservoir prediction.Traditional reservoir parameter prediction methods based on linear hypothesis,such as sparse pulse inversion,have been unable to reach the requirements of fine characterization of the reservoirs.Therefore,it is necessary to establish a non-linear reservoir prediction method to solve the problem of complex reservoir characterization.As one of the fastest growing branches of artificial intelligence,machine learning can mine features and relationships hidden in large data sets through autonomous learning.So it can achieve fully mine the lithology and physical information contained in seismic data by using machine learning in reservoir prediction,so as to achieve high-precision reservoir parameter prediction.In the current reservoir prediction methods based on machine learning,the influence of seismic facies on reservoir is offen neglected,which will undoubtedly reduce the accuracy of prediction results.In this paper,a method of reservoir parameter prediction under seismic facies control is proposed.First,spectral clustering method is used to divide seismic facies,and then the seismic facies information is added into reservoir parameter prediction process as a constraint condition.In the selection of input characteristic parameters of spectral clustering,we choode cepstrum characteristic parameters of seismic data to replace conventional seismic attributes.It effectively solves the problems of unclear correspondence between seismic attributes and geological characteristics and information redundancy.Experiments show that the cepstrum characteristic parameters of seismic data can fully excavate the hidden geological features in seismic data and improve the accuracy of seismic facies divisionOn the basis of seismic facies division,the convolution and fully connected neural network models are established respectively to predict reservoir parameters.In traditional neural networks,the connection between different nodes is inner product operation,which is not conducive to fully mining hidden information and suppressing noise in data.In this paper,the convolution operation is introduced into the neural network structure,and the convolution operation is combined with the inner product operation.Based on the structure of convolution and inner product neural network,an optimization algorithm of network parameter operator is presented and applied to reservoir parameter prediction process.This method is completely driven by seismic data and is not restricted by convolution model,so it can obtain high precision reservoir prediction results.Machine learning requires massive data for training,but in actual seismic exploration,due to economic costs,exploration and development degree and technical conditions,the number of samples that can be used for machine learning is often limited.In this paper,virtual sample generation theory is used to reservoir prediction process,and comprehensively uses the methods of extreme learning machine,data trend estimation and hypersphere characteristic equation to solve the small sample problem.The application of actual data shows that using virtual samples and small samples to build predict model together can effectively reduce the randomness of small sample modeling and approach the modeling results of sufficient samples to the maximum extent.