Identification And Prediction Methods For The Carboniferous Volcanic High-quality Reservoir In Chepaizi Bulge Of Junggar Basin,Northwestern China

Carboniferous volcanic reservoirs are characterized by the complex origin and the high reservoir non-homogeneity.The reservoir quality impacts the production of oil&gas a lot.However,the complicated subsurface not only severely restricts the data quality of geological and geophysical measurements,but also makes it difficult for the traditional semi-quantitative and quantitative reservoir characterization methods to provide satisfactory reservoir identification and prediction results.To address such problems,this thesis takes the Carboniferous volcanic reservoirs in Pai 66 block in Chepaizi bulge of Junggar basin as the research target,makes full use of geological and geophysical data such as cores,logs and seismic,and proposes a complete set of reservoir characterization and prediction methods based on geostatistics and deep learning methods with geological theory for volcanic reservoirs.It not only improves the prediction accuracy of high quality reservoirs in Carboniferous volcanic rocks,but also reduces the cost of manual interpretation and identification in the reservoir prediction process and improves the work timeliness.The Carboniferous volcanic reservoir located in the northwest margin of Junggar Basin is the rare proven world-class large-scale Paleozoic volcanic reservoir.Its reservoir is mainly concentrated in the weathered crust of volcanic rocks at the top of the Carboniferous formation.In this thesis,by using geological data such as cores and thin sections,the factors influencing reservoir development and distribution of Carboniferous volcanic reservoirs are firstly summarized into three categories: lithology,dissolution,and fracture.Lithology,as a comprehensive carrier of many geological factors,becomes one of the important influencing factors.Accurate lithology identification is a necessary condition for quality reservoir prediction.This paper firstly identifies the lithologies in the work area as: andesite,volcanic breccia,tuff and basalt through core data analysis,and proposes the logging lithology identification and prediction models which based on Bayesian statistics method and deep learning method to take the information of vertical spatial dependencies of geophysical data and continuity information of volcanic rocks into account.Compared with conventional machine learning methods and traditional linear identification methods,the lithology identification results provided in this paper have been improved in terms of accuracy and algorithm efficiency.The average probability score has increased from 0.509 to 0.676.Dissolution is another important factor influencing the reservoir of Carboniferous volcanic rocks.In this paper,the reservoir space of volcanic reservoirs in Block 66 was manually identified as dissolution pores,holes and fractures according to the cores and imaging logging data.Based on the semi-quantitative analysis,the reservoir statistics parameters were extracted and corrected by processing the imaging log pseudo-images of12 wells in the study area.An artificial logging curve named dissolution comprehensive value(DCV)is generated by taking geometric average of the statistics parameters.We believe that this artificial reservoir characteristic indicates the dissolution degree of volcanic rock well.A deep neural network(DNN)model was used to perform a joint well-seismic inversion to map the DCV curves to the full-zone 3D space.The 3D inversion results are validated by the blind well test with the accuracy as 0.836.The prediction is also consistent with the priori geological perception and the dissolution pattern of the volcanic weathering crust.To address the problem of the poor reservoir prediction provided by common methods,this paper proposes the Convolutional Long-short Term Memory(Conv LSTM)model to identify and predict high-quality reservoirs in the volcanic weathering crust with the post-stack seismic data and the 3D DCV inversion result.The Convolutional Longshort Term Memory model combines the advantages of the Convolutional Neural Network and the Recurrent Neural Network so that it can effectively extracts the spatial dependencies information and reservoir continuity information among the measurement points in the 3D sample.The 3D high-quality reservoir prediction provides by the proposed method gets high accuracy and keeps consistent with geological knowledge.The average F-measure of the blind test data is increased from 0.583 to 0.668.This paper develops a complete set of volcanic reservoir characterization and prediction method based on geostatistics and artificial intelligence and achieves great results in the Carboniferous volcanic weathered crust of Pai 66 block,Chunfeng oilfield.This paper demonstrates the feasibility and value of the geostatistics and artificial intelligence methods in practical applications.The proposed cross-disciplinary integrated research method based on mathematics,statistics,computer science and earth science has broad application prospects.