Research On Intelligent Lithology Identification Method Based On Invariant Information Representation Of Reservoir Meta-Targets With Logging Curves

As conventional oil and gas reservoirs gradually become depleted,unconventional oil and gas reservoirs have become the main force for the development of most mature oilfields.Therefore,the development of unconventional oil and gas reservoirs is crucial for achieving long-term stable production growth of oil and gas resources.The use of geophysical logging curves for lithology identification is a crucial step in accurately grasping the geological properties of target oil and gas reservoirs and achieving precise positioning of sweet spots.However,due to the frequent river channel transformations and multiple periods of river channel misalignments,the complex changes in vertical and horizontal directions of target sand bodies and strong heterogeneity of reservoirs,the identification of single and thin sand bodies is difficult.On the other hand,traditional lithology identification techniques based on classical machine learning methods such as neural networks cannot effectively incorporate geological priors,and the input features used generally lack well-to-well invariance,limiting the extrapolation ability of the machine learning models.Therefore,based on statistical computing,signal processing and characterization,pattern recognition,and machine learning techniques,this study aims to effectively enhance the capability of representing geological information and identification performance of unconventional reservoirs by focusing on the intelligent lithology identification method based on invariant information representation of meta-targets with logging curves.Firstly,in order to effectively construct meta-targets of reservoirs,it is necessary to use well logging curves to automatic layering of reservoirs.However,existing methods for automatic logging curve stratification do not make full use of the cross-correlation information between heterogeneous logging curves,and have weak structural description capabilities for local homogeneity of reservoirs,making it difficult to achieve reliable representation of metatargets of reservoirs.Therefore,relevant differential and structural tensor features with structural invariance information representation capabilities for local homogeneity of reservoirs are proposed.Based on this,a clustering-based automatic stratification method for logging curve integration is proposed to accurately and reliably construct reservoir meta-targets(automatic stratification).Experiments on actual data from shallow tight oil reservoirs and shale oil reservoirs in the Daqing Oilfield show that the proposed method can achieve fine automatic stratification of logging curves for unconventional oil and gas reservoirs with different geological properties,effectively realizing the preliminary construction of meta-targets,with great application potential.In addition,in order to effectively identify the lithology of reservoir meta-targets,considering that existing feature extraction methods for lithology have not fully mined the local texture invariance information of logging curves.Therefore,two feature description methods with the ability to describe local texture invariance information of logging curves are proposed:the logging curve local texture invariance features composed of multi-view Local Binary Pattern(LBP)features and Hu invariant moments features,which effectively represent the lithology-related information of reservoir meta-targets.Specifically,considering the intrinsic multiscale characteristics of logging data,a method for identifying reservoir lithology based on multi-scale invariant feature fusion of meta-targets was realized by adopting a multiscale invariant combination feature based on the multiscale logging curve local texture invariance features and structure tensor invariance features as inputs and using typical machine learning algorithms as the lithology classifier.Through comparative experiments on a real multi-well dataset in the Daqing Oilfield for lithology recognition,the experimental results show that the proposed multiscale fusion feature have stronger ability to describe lithology of meta-targets than the original logging curve data,and can achieve better lithology recognition results,providing a new idea for lithology information representation in existing logging curve lithology recognition and related technologies.