Logging Curve Prediction And Reservoir Identification Based On Deep Learning

Petroleum exploration is critical to the economic development and national production.At present,reservoirs that can be explored easily in China have been almost fully developed.What can be developed in the future are of low resistivity,low porosity and low permeability.Existing analysis technologies cannot mine accurate enough reservoir information from the logging big data.The inaccuracy of the mined reservoir information has limited the large-area exploration.The logging curve is a kind of geophysical parameter which changes along with depth of the wellbore at the same interval.It contains abundant geological information and has high vertical resolution.Fully mining the knowledge and rules contained in the logging curve can give us a deeper understanding of the formation properties.But in the process of practical exploration,due to reasons of geological conditions,the cost budget,limited number of drilling,only parts of the study area can be drilled.The measured data is not enough to fully understand the nature of the formation in the study area,resulting in difficulties of deep reservoir identification in some special circumstances,and finally resulting in a waste of non-renewable resources.To this end,this thesis deeply explores the issues of logging curve prediction and reservoir identification based on deep learning.The main research content is as follows.Firstly,the depth control curve method is used to calibrate the logging curve.The real logging curve is obtained by effectively filtering the noise signals in the raw logging data with wavelet decomposition.By properly pretreating the logging curve,the influence of various random interference and non-formation factors on the logging curve can be reduced.The obtained logging curve can truly reflect the nature of formation and pore fluid,provide real and complete data support for the subsequent model training.Secondly,a logging curve prediction method based on the generative adversarial network(GAN)is proposed It mainly includes two parts:the generation model and the discrimination model of the logging data.By adding constraints,the distribution law of the real logging data in the study area is learned,and the mean square error is introduced into the objective function of the original GAN.The purpose is to improve the learning ability of the model and the accuracy of curve prediction.Prediction experiments of the logging curve between wells and experiments of the logging curve in the missing section are conducted.The results showed that the proposed method has better prediction effect than Kriging and methods based on all connected neural network.Finally,a multi-scale reservoir identification method based on recurrent neural network(RNN)is proposed.In order to make full use of the spatial correlation between logging data,and to accurately identify reservoir layers with similar physical properties,an identification model of reservoir and non-reservoir in large scale is established based on RNN and the finer-scale identification of reservoir is realized by means of a series of multi-layer fully connected neural network(FCNN).This method not only considers the correlation between the logging data,but also distinguish similar layers in a multi-scale way.It solves the difficulties of feature extraction and improves the rate of layers recognition.The experimental results show that this method is practically effective.In summary,this thesis realizes the prediction of logging curve and reservoir identification based on deep learning.The achievements of this thesis are helpful for geologists to accurately recognize the study area,reduce the cost in the process of reservoir development,and improve the recovery rate of reservoir.