支持向量机在大庆齐家凹陷测井解释中的应用

Application of support vector machine in logging analysis of Qijia sag in Daqing Oilfield

(1.大庆石油学院,黑龙江大庆163318; 2.中国石油大学机电工程学院,北京102249; 3.中国石油集团大庆石油管理局钻井技术服务公司,黑龙江大庆163358)

Daqing Petroleum Institute, Daqing 163318, China

在大庆齐家 古龙凹陷葡西地区的含油层系中，薄差油层、低阻油层、高阻水层并存，储层的岩性、孔隙结构复杂多变，开发井的解释符合率较低，远不能满足油田高效开发的要求。为了提高测井解释精度，之前已经采用许多学习算法（如Bayes方法）对该地区的油水层进行划分，但都未能达到理想的效果。为此，以大庆齐家凹陷某口井的测井资料为例，探讨了支持向量机方法在油气识别中的应用。支持向量机方法基于统计学习理论，具有全局优化、泛化能力强等优点，适合对不同模式进行分类，在建立分类模型时仅依赖原始测井数据，无需完全依赖地区经验公式或经验数据，因此减小了由经验带来的误差。利用一口井的测井资料，提取了51层样本作为测试参数，分为两组：一组取27层为训练样本，24层为测试样本；另一组取20层为训练样本，31层为测试样本。分别采用支持向量机方法和Bayes方法进行了油气预测，在两种训练样本条件下，支持向量机方法的预测精度分别达到了100%和93.5484%，而Bayes方法的预测精度分别为77.7778%和77.4193%，这表明利用支持向量机方法对未知油气层的属性进行正确识别是可行的。

In Puxi area of Qijia-Gulong sag in Daqing oilfield, thin-poor oil layers, low-resistivity oil layers, and high-resistivity water layers coexist. The lithology and the pore structure are complex, which result in a low interpretation coincidence rate for development wells and cannot satisfy the requirements for effective development. To improve the interpretation accuracy, although many methods (such as the Bayes) have been applied to classify the oil/water layers in this area, the results are not ideal. By taking the logging data of one well in Qijia sag at Daqing oilfield as an example, the application of the Support Vector Machine (SVM) in hydrocarbon recognition was probed. Based on statistical learning theory; the SVM is very good in global optimization and generalization and is suitable for classifying different patterns. The establishment of classification model merely depends on the original logging data does not completely depend on the empirical formula or empirical data of the area, which may decrease the error caused by experience. Layer samples were extracted as the test parameters from the logging data of one well and were divided into two groups, each with 51 samples. For the first group, 27 layers were taken as training samples and the other 24 as testing samples; for the second group, 20 layers were training samples and the other 31 were testing samples. Both the SVM and the Bayes were applied respectively in hydrocarbon prediction, indicating that the prediction accuracy of the SVM was 100% and 93.5484% for the two groups, which were higher than that of the Bayes (77.7778% and 774193%). Therefore, the SVM is feasible in correctly recognizing the properties of uncertain hydrocarbon zones.