Comprehensive Interpretation Of Tight Sandstone Gas Reservoir In DJ Block,Eastern Ordos

The upper Paleozoic tight sandstone gas reservoir in DJ block in the eastern margin of Ordos Basin has undergone compaction,compaction and dissolution of unstable components,and is now in the late diagenetic stage B,with low porosity,poor permeability and strong heterogeneity,and it is difficult to identify the reservoir fine logging evaluation and sweet spot area.Therefore,a high-resolution sequence stratigraphic framework of the study area was divided and constructed by using outcrop,core thin sections and related test and logging data based on sequence stratigraphy theory and comprehensive correlation interpretation method of single well and connected well.On this basis,the sedimentary facies of 32 4-order sequences in benxi formation-Shihezi Formation are further explained and their evolution characteristics are analyzed.Using more than 800 groups of measured porosity data(640 for training samples,160 for test samples),5 groups of 32 sample data sets were determined by logging parameter traversal combination strategy.The porosity prediction model of 32 types of different sample attribute data sets was tested by using the four-layer BP neural network algorithm with five times of cross-validation.The optimization of logging parameters and over-fitting of model in machine learning prediction of tight sandstone porosity were studied.Based on the optimized porosity model of tight sandstone,the porosity plane distribution characteristics of four sequences in shanxi Formation are explained.The main findings were as follows:1.Based on the data of 43 CBM Wells and tight sandstone gas Wells,the strata of Benxi to Shihezi Formation of Upper Paleozoic in the block are divided into 14 tertiary sequences.There are two in benxi group,SQ1 ～ SQ2;Two in Taiyuan Formation,SQ3 ～ SQ4;Two in Shanxi Group,SQ5 ～ SQ6;Lower stone box group,SQ7 ～ SQ10;Two upper stone boxes,SQ11 ～ SQ14.Thirty-two quaternary sequences were divided,including four in benxi Formation,PS1-PS4;4 in Taiyuan Formation,PS5-PS8;There were 5 in Shanxi group,PS9-PS13;10 lower stone boxes,PS14 ～ PS23;There are 9 upper stone boxes,PS24 ～ PS32.2.The Benxi and Taiyuan formations in the study area are barrier lagoon-carbonate platform complex systems,the Shanxi Formation is delta system,and the Shihezi Formation is river system.The main reservoir sections of tight sandstone gas in this area,The 2nd member of Shan Formation(PS9 ～ PS11),show a complete transgression/regression cycle.The thickness of water-inrush and water-regressive sequences are similar,and the waterinrush and water-regressive sequences are slow and the amplitude is not large.The underwater branch channel,mouth bar and mat sand facies are relatively developed.The shan-1 member(PS12 ～ PS13)also developed a complete transgression/regression cycle,showing the characteristics of rapid water inflow and rapid water regression.The waterinflow sequence is thick but fine-grained,mainly composed of mudstone deposits in interdistributary bay or pre-delta front.3.The combination of parameter traversal and cross validation can effectively solve the problem of parameter optimization strategy and over-fitting in the MACHINE learning model of DJ block porosity.For the prediction model of tight sandstone porosity by BP neural network,the selection and combination of logging parameters are critical.The prediction accuracy of porosity is positively correlated with the number of attributes of training data.With the increase of the number of attribute parameter combinations,the accuracy of BPNN model is improved.There are significant differences in porosity prediction error levels among the log parameter combinations representing different lithology and physical properties.Among 32 models,the maximum relative error is 38%,the minimum is 24%,and the maximum difference is more than 14%.Four kinds of conventional logging parameter curves,such as acoustic time difference,compensated neutron and natural gamma,show obvious positive effects when they are input as sample attributes.The prediction accuracy of the full parameter model is the highest.