Research Of Water Injection Profile Prediction Based On A Data Driven Method With Few Samples

The areal and vertical heterogeneity of onshore oilfields in China is extremely serious.The planar and interlayer displacement difference of the reservoir is further enlarged,and the development has become increasingly difficult after long-term water injection development.An accurate understanding of water absorption difference between layers is the basis of improving the effectiveness of water flooding development.Therefore,the prediction of water injection profile is an important task in mature oilfields with high water cut,and it is the guarantee for the efficient development of mature oilfields.Field test of water injection profile is a traditional way to grasp the water absorption of layers,but this method is costly,time-consuming,and difficult to conduct long-term monitoring.How to use the limited water injection profile data accumulated during the reservoir production period to develop an effective water injection profile prediction model is still an important engineering issue.To overcome the problem of insufficient water injection profile in oilfields and difficulty grasping the water absorption of the reservoir during the full production period,this dissertation proposes a set of water injection profile prediction methods based on data driven with few samples for the first time.Meanwhile,the main attributes affecting the water injection profile can be identified and the prediction of water injection profile in the full production period is realized.Firstly,an inter-well connectivity calculation method（MD-SDTW）based on the similarity of multi-dimensional time series is proposed by modifying the traditional dynamic time warping algorithm（DTW）.The gradient of the time series is introduced to eliminate the influence of the singular problem of DTW on the accuracy of similarity analysis.The Mahalanobis distance is adopted to measure the local distance of multi-dimensional variables to improve the accuracy of the similarity analysis of the input-response signals.Taking the daily water injection rate and injection-production ratio as the input signal of injection wells,and taking the daily liquid production rate and water cut as the response signal of production wells.A multi-dimensional time series is then constructed to analyze the connectivity between injection and production wells.The connectivity analysis was conducted on the 22X well block,and the inter-well connectivity was consistent with the tracer monitoring results,which verified the proposed approach.It lays the foundation for the fusion of water injection profiles.Then,a semi-supervised feature selection method（GLS²FS²）based on graph Laplacian theory andL_2,1 regularization has been developed.The graph Laplacian theory is used to construct a neighbor graph using the labeled and unlabeled data,obtaining a semi-supervised scatter matrix that preserves the spatial structure of the data.The L_2,1 regularization is adopted to ensure the sparsity of the feature transformation matrix,and establish the objective function of GLS²FS².The transformation matrix then can be solved iteratively,and the importance of each feature is evaluated based on|_iW|₂.It is verified that the GLS²FS² can effectively eliminate the redundant features and improve the prediction accuracy of the model through the test in regression datasets.The feature selection test of water injection profile based on GLS²FS² was performed in the 22X well block,and 25relevant features such as porosity,permeability,and injection pressure were selected from 43attributes for modeling.Lastly,a semi-supervised regression method based on XGBoost Co-training（En XGB-Co Reg）is developed to predict the water injection profile for the well block with a small amount of water injection profile.Multiple training sets are generated by bootstrapping,and multiple XGBoost models are initialized by genetic algorithm to enhance the diversity of base learners in the co-training paradigm.The high confidence samples selected by the model are gradually added to its own training set to ensure that the prediction performance of the model can be steadily improved during the co-training.It avoids the risk of performance degradation in traditional semi-supervised learning methods.Additionally,a safe and reliable water injection profile prediction model is established.The regression datasets were used to test the advantages of the En XGB-Co Reg algorithm in few shot learning paradigm.The results show that its prediction accuracy is higher than that of traditional machine learning methods.The practical application of the En XGB-Co Reg algorithm in the22X well block shows that the prediction performance of the En XGB-Co Reg algorithm is better than that of the KH splitting method and other supervised/semi-supervised learning methods such as SVM and Co Reg.It can accurately excavate the variation law of the water injection profile.Furthermore,a transfer learning method based on joint domain adaption and ensemble learning（JDA-Ens）has been developed for water injection profile prediction of well block without water injection profile.JDA is used to jointly adapt the marginal distribution and conditional probability distribution of the source and target domains to mine the common feature representation between the domains.The monitored profile of the source well block then can be used to assist in establishing the predictive model of the target well block.In order to improve the learning performance of the JDA algorithm for high-dimensional characteristic data,the ensemble model is built for providing pseudo labels for the unlabeled data and constructing the final predictive model.Meanwhile,to improve the prediction accuracy of the ensemble model and reduce the work of manual parameter tuning,a parameter tuning strategy based on the combination of genetic algorithm and cross-validation is established.And the JDA-Ens algorithm was tested and verified in well22NA-11 and well 25E-2.Summing up,a prediction method for water injection profile prediction under few samples and zero samples is developed in this thesis,which lays a theoretical foundation for the efficient development of mature oilfields with high water cut.