Reservoir Characterization Prediction From Well Logs

After 40 years' water flooding, Daqing Oilfield has stepped into the later stage of high water-cut. At this stage the inner structure of reservoir is extremely complicated and the oil is seriously separated, which require the study of the reservoir fine description to determine the water-flooded status and the distribution of remaining oil. The determination of the reservoir characterization, which serves as basic parameter in the development of oilfield, is one of the most important tasks of the reservoir fine interpretation. At present the most economical and convenient method to predict the reservoir characterization is using well log data. However the present log interpretation model cannot satisfy the need of the fine description, therefore log interpretation models with higher precision are needed.A method of using conventional well logs to predict reservoir characterization such! as porosity, permeability and present water saturation is introduced in this study. This method breaks through the limit of conventional log interpretation methods, make the best use of available well logs and increase the precision of log interpretation largely.The steps of this method include the classification of wells, the classification of reservoir, the pretreatment of log data, the selection of log variables and the establishment of regression model. Through the classification of wells by drilling time and log tools and the classification of reservoir by oil formation, the heterogeneity of the whole well can be transformed to the relatively homogeneity of every class. In such class, one reasonable and effective model can be built to describe the characterization of the whole class. The pretreatment of log data calculates corresponding log values for every core sample to solve the problem that the log data and the core data are unmatchable and reduce the influence of error caused by the jog process. All possible subsets regression selects 5~6 sensitive log variables from 11 available log variables to consist of the sensitive log models by comparing the absolute error and the relative error of different regression models. Power transformation is used to build up non-linear models based on the sensitive log models.The final regression models are non-linear models comprising the best power transformation of 5～6 log variables. The absolute and relative errors of the final regression are much less than those of the original log interpretation. The purpose of increasing the precision of log interpretation models is achieved successfully.