Research On Intelligent Modeling Method Of Lithological Profiles Under Complex Geological Conditions

The lithology profile is a key component in studying the spatial structure and morphology of rock strata,which is of great significance for reservoir description and characterization.The lithology profile modeling is usually divided into two steps.First,logging data are used to predict the lithologic sequence of the strata.Then,analysis is conducted on the horizontal and vertical correlation of the lithology of the strata.In traditional methods,this process is completed by geologists based on experience and geological knowledge.However,as logging data continues to accumulate and exploration targets towards unconventional oil and gas reservoirs,traditional methods have become increasingly constrained,with the emergence of multiple solutions and an efficiency that is difficult to meet demand.Therefore,how to efficiently and accurately construct lithology profiles has become a key issue in the new era of oil and gas exploration.In recent years,machine learning technology has achieved rapid development in many fields.Its powerful data mining and feature extraction capabilities provide new ideas for solving the bottleneck in the traditional modeling method of lithologic profiles and have received widespread attention.However,unconventional oil and gas reservoirs have strong heterogeneity and complex geological environments.The nonlinearity between reservoir parameters and logging responses has made intelligent modeling of lithologic profiles face some challenges:(1)Due to differences in geological structures,sedimentary environments,and drilling processes,there may be discrepancies in the distribution of logging data among different wells;(2)Core sampling and logging are costly,making it difficult to obtain reliable lithology labels by frequent core sampling operations;(3)The mathematical modeling task of constructing lithology profiles is complex,and the predicted lithology profiles need to conform to the objective geological distribution law.To address the aforementioned challenges,this dissertation conducts research on intelligent modeling methods for lithology profiles under complex geological conditions.Based on the lithology profile modeling process and the characteristics of actual working conditions in oil exploration,the research is divided into three parts:construction of a lithology identification model oriented towards the entire working area,construction of a lithology identification model oriented towards a single well,and construction of lithology profile.The specific research content can be concluded as follows:(1)Construction method of domain-generalized lithology identification baseline model based on contrastive difference:to address the issue of distribution discrepancy in the logging data among different wells during the construction of baseline models,a domain-generalized approach based on contrastive difference is proposed for constructing lithology identification baseline models.By leveraging the logging data and logging information from existing wells in the well group and incorporating domain generalization ideas,the lithology identification baseline model is constructed.First,domain invariant features are learned through adversarial autoencoders.Specifically,the shared encoder among different wells is used to extract the high-level feature representations.Meanwhile,contrast difference constraints are imposed on the feature representations to align samples with the same lithology and separate samples with different lithology from different wells.Additionally,an adversarial learning approach is introduced to make the learned features more robust.Finally,a classification layer is added,and label information is introduced to learn the features which are both discriminative and domain invariant.Experiment results on a real dataset demonstrate that the proposed method can effectively reduce the differences in logging data between different well positions without introducing target domain information,and can extract their common features,obtaining considerable lithology recognition accuracy.(2)Construction method of lithology identification single well model based on multi-scale active adaptation:to address the issue of high core logging costs in new wells,a single-well lithology identification model based on multi-scale active adaptation is proposed.By combining active learning and domain adaptation ideas,the baseline model is optimized to establish a lithology identification single well model that meets the requirements of exploration accuracy.First,the discriminator is used to select samples similar to the target domain in the source domain to perform initial training on the model.Then,multi-scale parameter adjustment is performed on the pre-trained model in coarse and fine-grained manners,in order to enrich the the number of target labels and improve the performance of the model.During coarse adjustment process,a clustering algorithm is used to select a small number of samples for querying real labels based on the distribution characteristics of logging data.Subsequently,a graph-based semi-supervised algorithm is used for label propagation based on similarity in the feature space,while retaining pseudo-labels with high confidence.Further,the selected pseudo-labels are detected,and unreliable pseudo-labels are removed.The model is optimized using labeled samples and corrected pseudo-labeled samples in the target domain.During fine adjustment process,two active learning algorithms are combined to select informative samples for label querying,and the model is further optimized using their real labels to improve the accuracy of the lithology identification single well model.The experimental results on a real dataset show that the proposed method outperforms multiple existing methods in different domains and can improve the accuracy of the lithology identification model with as few target domain lithology labels queried as possible.(3)Intelligent modeling method for lithology profile based on geological constraints:to address the complexity of mathematical modeling in constructing lithological profiles,an intelligent modeling approach that incorporates geological constraints is presented.The problem of predicting lithology profiles is formulated as a determination problem of potential sand body connections between adjacent wells,and pruning is introduced to achieve lithology profile construction.Firstly,the wavelet transform is applied to the single well logging curve to identify the abrupt points or areas between frequency structural segments for stratigraphic division of the single well.Then,using the marker layers in the stratigraphic division results as starting points,the adjacent sand bodies with similar local logging shapes and overall logging trends in different wells are connected using convolutional modules at the front end.At the back end,a correction algorithm is designed based on the law of geological extension to prune the connected results.This step enables the acquisition of the final correlation results for sand bodies,which are then used to construct the lithology profile.The experimental results on real datasets demonstrate that the method proposed in this research produces lithology profile results that are well-matched with manually constructed results by geologists and effectively integrate geological constraints and expert knowledge.Based on the aforementioned research,this dissertation establishes a research framework for "baseline model construction for lithology identification-single-well lithology identification model optimization-lithology profile construction",and constructs a methodology for intelligent modeling of lithology profiles,providing a feasible technical approach to accurately characterize the geological characteristics of reservoirs and improve the efficiency of oil and gas exploration.