Study On Prediction Method Of Reservoir Parameter Based On Rough Set Theory And Support Vector Machine

Quantitative characterization of reservoir characteristics is one of the most basicworks in revealing the law of reservoir oil and gas movement, selecting reasonabledevelopment layer series, and guiding the development of production of oil and gas fieldto enhance oil recovery efficiency. Experiences of exploration and development of oiland gas show that there are closely nonlinear relationships between logging curves andreservoir characteristics. The same result resists in seismic attributes and reservoircharacteristics. With the development of oil and gas exploration and development, thetraditional reservoir parameter prediction method based on linear assumption has beenunable to meet the requirements of fine description of reservoir characteristics. Thesupport vector machine is a new machine learning algorithm suitable for solvingnonlinear problems. Now support vector machine has been a hotspot in research onnonlinear intelligent method. Rough set theory is a kind of tool for dealing withimprecise, uncertain and incomplete information data. Recently, it has been widelyapplied in many fields of science and engineering. Reservoir parameter modeling ofrough set theory and support vector machine are constructed based on logging curves andseismic attribute data in two work areas in this paper, striving for exploring a new way ofreservoir parameter prediction and simulation. The main results are as follows:（1）The research status of reservoir parameter prediction method, support vectormachine in prediction of reservoir parameters and rough set theory are reviewed. The keyof the problem in the application of support vector machine in reservoir parameterprediction, and the research direction in the future are improved.（2）The most widely used radial basis function is selected as the kernel function.The algorithms flow of three methods of kernel function parameter optimization, that isgrid search method, genetic algorithm and particle swarm optimization are analyzed,which lays important foundation for reservoir parameter prediction modeling based onsupport vector machine method.（3）The basic concept, attribute reduction mechanism and method of rough settheory are introduced. The logging curves and seismic attributes in two research workareas are reduced using rough set method. The most closely groups of logging curves andseismic attributes with reservoir parameters are selected. （4）As the input ofsupport vector machine, the selected loggingcurves are used toconstruct a reservoir parameter prediction model. In kernel function parameteroptimization, grid search method, adaptive particle swarm optimization and adaptivegenetic algorithm are adopted respectively. The applied results of model prediction intwo work areas show that the prediction model can predict the porosity and permeabilitywell, and the prediction results are better than that of conventional log interpretationmethod. There are different in precision and time consuming between models based ondifferent kernel function parameter optimization. The prediction precision by usinggenetic algorithm and particle swarm optimization are higher than that of grid searchmethod. It consumes the least training time by using grid search method. The predictionaccuracy for reservoir permeability is lower than that of porosity using the constructedsupport vector machine model.（5）As the input of support vector machine, the selected seismic attributes are usedto construct a reservoir parameter prediction model. The applied results of modelprediction in two work groups of Toutunhe and Xishanyao show that the predictionmodel can predict the porosity and sandstone thickness well, and the prediction resultsare better than those of BP neural network method and multivariate linear regressionmethod. The planar prediction diagram shows that the prediction result of support vectormachine is more in line with the sand body distribution and physical characteristic ofsedimentary facies. Among three parameter optimization methods, the prediction resultof using particle swarm optimization is the best. It consumes more time in predictingsandstone thickness than porosity. The prediction accuracy for porosity is higher than forsandstone thickness.