Research On Heterogeneous Reservoir Data Inversion Method Based On Deep Learning

Reservoir inversion is an important content in petroleum,natural gas and other solid mineral exploration and development.It’s a basic link to extract reservoir information hidden in the geophysical logging data.With the development of artificial intelligence,machine learning,especially deep learning has penetrated into all aspects of petroleum exploration and development.It has become a hot technology in reservoir inversion research.Due to the complex sedimentary environment and tectonic movement,lithologic reservoirs have the characteristics of strong reservoir heterogeneity and complex pore structure,resulting in the strong nonlinear relationship between the reservoir attributes.Data-driven deep learning can deeply mine the nonlinear implicit information between the attributes,which can intelligently invert reservoir data according to the relations between existing reservoir attributes,providing basic data support for further geological research and scientific decision-making.Accurate and intelligent inversion of heterogeneous reservoir data is of great engineering significance for accelerating energy exploration and development.Based on deep learning technology,this paper takes carbonate rock,shale and other heterogeneous reservoirs as the research objects,and integrates multi-disciplinary methods to conduct quantitative characterization of reservoir logging data and intelligent inversion of lithology spatial distribution.Meanwhile,focusing on the key scientific issues of uncertainties,unidirectional application and Gaussian hypothesis in the combination of deep learning and logging engineering fields.The main research contents are as follows.A heterogeneous reservoir logging data inversion method is proposed based on deep learning and attention mechanism to solve the problems of poor quality and incomplete logging data.Firstly,a logging feature attention module is constructed to focus on the spatial correlation between the characterization target and related logging attributes at the same depth,which can extract logging spatial features autonomously.Then,a bidirectional GRU network is designed to capture the logging depth features of the same logging attribute in a certain depth range,and realize the intelligent inversion of logging data and quantitative identification of lithology.It provides fast and effective data for further geological research and reservoir modeling,reduces logging cost,improves drilling and completion strategy.Finally,the method is applied to the carbonate reservoir of Mishrif formation in Iraq,and evaluates the performance from model accuracy and model fitting degree.On the basis of further address the data quality and data integrity issues,a heterogeneous reservoir logging data inversion and uncertainty analysis method is proposed to solve the problem of the uncertainties in heterogeneous reservoir inversion system.Firstly,studying the characteristics of logging attributes in space,depth and time dimensions,a multi-stage attention method for extracting logging spatial features,logging depth features and logging time features is proposed to realize the accuracy and intelligent inversion of reservoir logging data.Then,a GRU＿Bayes network framework is constructed by modeling the distribution of deep neural network weights based on Bayesian theory,which quantitatively captures the data uncertainty and model uncertainty of the inversion system.It can achieve system risk assessment,assist the system to make better decisions,and improve the reliability and engineering applicability of the reservoir inversion system.Finally,take the acoustic logging inversion in carbonate reservoir as an example,the performance of the method is verified and evaluated from the perspective of accuracy,uncertainty and scalability.Under the geological constraints,a reservoir lithology spatial distribution inversion method is proposed to address the problems of the computational cost in the spatial distribution inversion and the unidirectional application in deep learning modeling.Firstly,the limitations of Kriging technique in the lithology spatial distribution inversion are deeply analyzed,and the finite weighted sum of the basis function is solved to approximate the Kriging spatial correlation process.A spatial-dependent GRU＿Kriging network framework is constructed to establish a direct connection between spatial coordinates and deep neural network,which can obtain more spatial features with less computational cost and memory complexity.Secondly,the weight coefficients of geological constraints are extracted based on Kriging technique,and the deep neural network of geological constraints is constructed.Under the constraints of domain knowledge,it realizes intelligent inversion of lithology spatial distribution.Then,quantitatively capture the uncertainty of the results to realize the reliability evaluation of the reservoir inversion system.Finally,the method is applied to shale reservoir,and evaluated from the reservoir lithology spatial distribution inversion accuracy,model reliability and time performance.Under the non-Gaussian hypothesis,a reservoir lithology spatial distribution inversion and uncertainty analysis method is proposed to address the problem of the Gaussian hypothesis in the reservoir inversion and reliability analysis modeling.Firstly,analyzing the non-Gaussian characteristics of carbonate rock,shale and other heterogeneous reservoir lithology deeply,and adopting the symmetric logarithm ratio transformation method to fit the non Gaussian characteristics to extract the spatial characteristics.The spatial distribution of reservoir lithology is inverted intelligently.Secondly,the random and dynamic non-Gaussian characteristics of the deep neural network weight parameters are analyzed,and a Gaussian transformation iterative Ensemble Kalman method(GS＿IENKF)is proposed to assimilate the non-Gaussian data in real time.It is guaranteed that the data still follows the original spatial distribution in the real-time update process.The reliability of the inversion system is quantitatively evaluated under the non-Gaussian hypothesis.Finally,the method is applied to the laterite nickel ore heterogeneous reservoir,and good results are achieved in the inversion accuracy,model fitting degree,model reliability and time performance.