Study On Logging Evaluation Methods Of Igneous Reservoir Based On Ensemble Learning

Igneous oil and gas reservoirs are widely distributed in Mesozoic to Cenozoic onshore and offshore basins around the world,showing remarkable characteristics of global development.However,the highly inhomogeneous nature of igneous reservoirs leads to significant variations in their production capacity,which poses a great challenge for reservoir evaluation.Over the past two decades,ensemble learning has become an important technique in the field of machine learning.ensemble learning produces a more accurate and robust prediction by combining the predictions of multiple underlying models.With its excellent nonlinear mapping capability,this method has been widely used in various tasks and fields,and ensemble learning has shown potential for application in igneous reservoir evaluation work.In the qualitative evaluation of igneous reservoirs,the potential connection between lithology,lithofacies,and reservoir fluid classes is often ignored,and past researchers have focused only on the qualitative evaluation of a single geological attribute,which limits the accuracy and reliability of the evaluation to a certain extent.Although some researchers have considered the intrinsic connection between multiple geological attributes when using conventional means for qualitative evaluation of igneous rocks,a systematic and complete evaluation method has not yet been formed.Meanwhile,the category imbalance of the data set will have a negative impact on the results of reservoir evaluation,so the category imbalance of the core scale logging data needs to be solved urgently in the workflow of reservoir qualitative evaluation.Models based on rock pore structure occupy an important position when performing permeability calculations in igneous reservoirs.As a key parameter reflecting the rock pore connectivity and pore throat distribution,the capillary pressure curve is closely related to the fluid flow behaviour.Currently,the models for predicting rock permeability using the capillary pressure curve are mainly divided into two categories: models based on the seepage theory describing the electrical conductivity and models based on Poisson’s law and Purcell’s theory.However,the scale of petrophysical experiments is limited by exploration costs.Logging data show significant advantages in the evaluation of igneous reservoirs due to its strong continuity,large data scale,and comprehensive types.Although there have been successful cases of using logging data to predict permeability in homogeneous formations,the correlation between logging data and permeability is low in igneous reservoirs,which poses a challenge to the accurate prediction of permeability.Therefore,the development of more effective permeability logging prediction methods for igneous reservoirs is an important direction for current research.In this paper,for the Middle and Cenozoic igneous reservoirs in the eastern depression of the Liaohe Basin,we summarise the classification scheme,geological features and logging response characteristics of the igneous lithofacies and lithology in this area.It also discusses the application effect of quantitative and qualitative evaluation methods of igneous reservoirs based on integrated learning with the logging and petrophysical data of the eastern depression.The paper specifically includes the following aspects:Firstly,in terms of qualitative evaluation of logging,this paper starts from the qualitative evaluation of a single attribute,and proposes an integrated learning-based lithology identification method,which uses adaptive multi-objective group fusion to improve the class imbalance of the dataset,and it is centred on the particle swarm optimisation algorithm,where the minority class samples in the dataset are oversampled by the optimised synthetic minority class oversampling technique,and the majority class samples are selected by the swarm instance selection downsampling.In lithology identification,the Extreme Cascade Forest method is used as a classification algorithm,Extreme Cascade Forest is an integrated learning method that uses extreme gradient enhancement to replace random forests in the cascade layer of multigranularity scanning cascade forests,and a disassembly method is used to convert a multi-classification problem into a series of binary classification problems.The results show that the method has good performance and application results,and is a reliable method for lithology identification.Subsequently,this paper introduces multi-label learning into logging evaluation for the first time and uses two mainstream ideas of multi-label learning,respectively proposing a problem transformation-based approach and an algorithmic adaptationbased approach to adapt to qualitative evaluation scenarios with different labelling complexities,and also proposes a multi-view evaluation metric,which evaluates the classification effect in an all-round way from the label perspective,the sample perspective and the global perspective.Aiming at the low-labelling complexity scenario of igneous lithofacies and lithology identification,this paper proposes a multi-label integrated learning method using the idea of problem transformation.The method uses K-means to cluster the dataset and determines the target for oversampling by cluster imbalance index,and then uses SMOTE for oversampling.The method also uses optimised multi-granularity scanning cascade forests as a classification algorithm to convert multi-labelled,multicategorical data into a series of binary classification problems in order to simplify the algorithm complexity.For this high labelling complexity scenario of igneous lithofacies,lithology and reservoir fluid classes,this paper proposes a multi-label integrated learning method using algorithmic adaptation ideas.The method uses multi-label SMOTE to deal with the category imbalance problem of multi-label datasets,and uses multi-label deep forest as the classification algorithm.All the above methods have achieved good application results.Finally,in terms of permeability prediction,this paper proposes a permeability prediction method based on a priori constrained Bayesian gradient boosting.The method uses a feature engineering scheme based on capillary pressure data,logging and petrophysical formulations with a view to generating a large number of highly correlated features to improve prediction.The scheme consists of forward feature synthesis based on a priori constraints and depth feature synthesis constrained by physical formulae.During the prediction process,this paper proposes a gradient boosting integrated learning method based on tree-structured Bayesian optimisation,which employs a histogram decision tree as the basic algorithm to mitigate the overfitting situation due to feature engineering to increase the feature dimension.The results show that the method outperforms the comparison methods in model validation,while possessing better generalisation ability than the other methods.In actual wells in the study area,the method predicts results as expected and can be used as a reliable method for formation permeability prediction.