An Applied Study Of Support Vector Machines For Reservoir Parameters Prediction

There are many conventional reservoir parameters prediction methods, such as empirical formula, regression analysis and so on, but most of which are based on linear or single variable models, while the geological parameter having these characteristics like large sampling numbers, high dimensions, more randomness, uncertainty and so on. Therefore, conventional reservoir parameters prediction methods have already not satisfy the needs of reservoir production yet.In recent years, Support Vector Machine (SVM) has become the research focus in statistical theory. It is a method to build data model, by using the kernel function, which implicitly makes the data from low-dimensional space mapping into a high-dimensional feature space, thus the nonlinear problem can be solved as a linear problem. This kernel mapping model provides a unified framework for most of modeling frameworks, and its solution ultimately can be ascribed to a constrained convex quadratic programming problem, which largely resolved problems, like over-fitting learning, nonlinear, curse of dimensionality, local minima and so on.Combine the prominent advantages of SVM with the practical issues of oil geological exploration to analyse and research. This paper is based on the research and evaluation of the existing reservoir prediction methods, a complete set of nonlinear intelligent logging reservoir inversion technique is proposed, the Support Vector Regression (SVR) model is used for analyzing well-logging data to predict reservoir properties parameters (permeability and porosity). The whole process and the key technology for the reservoir parameter prediction is introduced in detail in this paper, including, using the cross plot and principal component analysis (PCA) for the well-logging sample selection, using particle swarm (PSO) for model parameter optimization, learning and describing the prediction model, and so on.Finally, according to the geological characteristics of the study area, the prediction model with a practical well-logging data is used to predict reservoir parameters for 5 wells in the study area, meanwhile comparing it with the prediction results of neural networks and core vector machines. The results show that the prediction model proposed in this paper is practicable, stable and effective for actual reservoir prediction, which can be,under a certain range and conditions,used for predicting other wells in the target oil area.