Inversion Of Parameters Of Fractured Vuggy Media Based On Machine Learning Considering Physical Process

The inversion of reservoir flow field parameters have important guiding significance for the formulation of reservoir well pattern adjustment,injection-production control and other measures.Compared with traditional methods such as well logging and well test interpretation,machine learning could mine the hidden relationships between data by analyzing and learning oilfield data.However,the current machine learning methods are still a kind of single data learning driven by pure data,and the physical interpretability of the results and the learning process needs to be improved;at the same time,the density of oil and gas development well patterns is generally small,and the well point data is still sparse training sample data.Even if the deep neural network is used,its prediction effect is still not good.In this regard,this paper adds the physical process information of oil and gas flow,that is,the percolation physics equation as a constraint condition,to the deep learning loss function item,establishes a new method of deep learning based on the physical process of oil and gas flow,and applies it to the study of fractured vuggy media.The main innovative work and conclusions are as follows.First,for the porous media model,sample data was formed based on the results of reservoir numerical simulation,and the inversion study of permeability parameters of single-phase flow and two-phase flow in porous media was carried out.The results show that: In the case of sufficient samples,both the traditional model and the new model in this article can obtain good learning and prediction results;as the data samples decrease,the prediction error of the traditional model also increases,but the new model in this article can still maintain high accuracy,even under the conditions of strong heterogeneity and injection-production relationship reversed.Subsequently,a fracture system was added on the porous media model,a strongly heterogeneous fractured media model was established.Data samples were obtained through numerical simulation of discrete fractures,and the inversion study of permeability parameters of single-phase flow and two-phase flow in the single-fracture and multi-fracture media was developed.The results show that: When the number of observation wells is small,due to the strong heterogeneity of the fractured media,the prediction results of pure data machine learning had huge differences from the actual situation,and the new method in this paper could identify the high permeability area in the reservoir,the error is less than half of the traditional model.Finally,on the basis of the fractured media model,the karst cave system is added to form a fractured vuggy media model.Use the Brinkman equation to characterize the fluid flow in the karst cave,and the Darcy-Brinkman equation is used as the physical process information of the fractured vuggy media flow.Use the numerical simulation result data of single-phase flow in fractured vuggy media as a sample for deep learning.The results show that: The traditional pure data machine learning has poor effect and cannot identify karst caves effectively.In contrast,this method can gradually determine that the middle part of the reservoir contains a certain high-permeability area,determine the general location of the karst cave,predict the presence of high-permeability bands outside the karst cave,and determine the karst cave and the location of the fractures,and can accurately predict its permeability parameters.