Obtaining variation function parameters in modeling geological attribute based on U-Net and CNN deep learning
In the modeling of geological attribute of oil and gas reservoirs,obtaining the variation function is es-pecially critical,which is generally obtained by fitting the experimental variation function to acquire parameters such as varve,azimuth,and abutment value.However,when the number of sample points in the study area is insufficient,it will lead to a poor fitting effect,thereby affecting the quality of attribute modeling.To overcome the shortcomings of traditional experimental variation function modeling and make the most use of spatial data,this paper proposes a new method based on U-Net and CNN networks to predict the parameters of the variation function.The data points extracted from the porosity plane model obtained by sequential Gaussian simulation are taken as the benchmark.Using the U-Net network structure,the porosity distribution is reconstructed to maintain spatial correlation.Subsequently,a CNN network structure is applied to the sample set for deep lear-ning,thereby developing a model to predict the variation function.The practical application shows that the prin-cipal range direction obtained by the proposed method in this paper deviates by only 1.52°from that obtained using the experimental range function,closely matching the distribution direction of sedimentary microfacies.Meanwhile,the obtained principal and secondary ranges closely align with the experimental variation function,confirming the reliability of the model's variation function results.At the same time,the method also simplifies the geological modeling workflow,reduces the subjectivity of finding the experimental variation function,and reduces the limitations posed by a small number of data points in the study area.It offers a novel approach for the predictive research of the variation function.
attribute modelingdeep learningmodel reconstructionSequential Gaussian Simulationvariation function
国家科技期刊平台
NSTL
万方数据
