支持向量机在多地质因素分析中的应用

Application of support vector machine to multi-geological-factor analysis

将支持向量机、人工神经网络、多元回归分析及参数乘积判别法4种算法分别应用于鄂尔多斯盆地塔巴庙地区40个致密砂岩储层的含气性评价,其预测结果与试气结果的平均相对误差绝对值分别为:0,4.63%,29.71%,18.75%。该实例表明:前两种非线性算法远比后两种线性算法优越;非线性算法中,支持向量机比人工神经网络优越;线性算法中,参数乘积判别法比多元回归分析优越。其根本原因在于:含气性与其相关地质因素(孔隙度、渗透率、含气饱和度)之间存在着复杂的非线性关系。因此,当描述一个研究目标与多个相关地质因素的复杂关系时,应提倡采用非线性算法,特别是在耗时巨大、多次反复进行多地质因素分析的数据处理作业中,应提倡采用支持向量机。因为它与人工神经网络相比,具有计算速度快、计算结果精度高的特点。另外,参数乘积判别法也具有简明、快速的优点,其精度远高于多元回归分析;而多元回归分析不仅计算速度快,而且还具有能表达研究目标与其相关地质因素之间亲疏关系的优点,可作为辅助手段。

Four different methods, including support vector machine （SVM）, artificial neural network （ANN）, multiple regression analysis （MRA） and parameter product decision （PPD）, were applied to the gassiness evaluation of forty gas-bearing layers in the tight sandstones of Tabamiao Area in Ordos Basin. Their mean absolute relative residual values between predicated results and gas test results are 0, 4. 63%, 29. 71% and 18.75%, respectively. This case study shows that the former two nonlinear methods （SVM, ANN） are very superior to the later two linear methods （MRA, PPD）. And the SVM is superior to ANN, while PPD is in turn superior to MRA. That is because there exists a complex and nonlinear relationship between gassiness and its related geological factors such as porosity, permeability and gas saturation. Therefore, ANN and SVM should be adopted to describe any complex relationship between the study target and its related multi-geological-factors. In particular, for time-consuming tasks of data processing with repetitious multi-geological-factor analysis, it is recommended that SVM should be used, because it is much faster and more precise than ANN. On the other hand, the case study also indicates that PPD has its advantages of conciseness and high speed. PPD has more precision than MRA, while MRA is fast in processing speed and can be used as an auxiliary tool to establish the order of dependence between the target and its related multi-geological-factors which cannot be estimated using the other three methods.