【作者】 马海；

【导师】 王延江；

【作者基本信息】 中国石油大学 ， 控制理论与控制工程， 2010， 博士

【摘要】 在石油勘探、开发领域中存在着大量复杂和不确定的因素,尤其在钻井过程中随时可能遇到不同的地质条件,如岩性的变化、可钻性的变化、地层压力的变化等等,这些都将给钻井过程带来复杂多变因素,如果不能准确了解和认识地下复杂的钻井环境,就会影响钻井施工,甚至造成钻井事故危及人身安全。论文对石油勘探开发中的多源信息进行了分析,重点对地层可钻性级值、地层压力、地层岩性等钻井地质特征参数的估计与预测方法进行了深入的研究。论文主要工作如下:对石油勘探开发中的信息来源及表达方式进行了研究,分析了它们与钻井地质特征参数的关系,构建了石油勘探开发的多维异构空间模型,将多源信息通过一个知识表达系统进行表征。对录井及测井信息与地层可钻性级值的关系进行分析,提出一种基于相关向量机的地层可钻性级值预测新方法,通过训练相关向量机,建立了地层可钻性级值预测的相关向量机模型。将量子粒子群优化算法与支持向量机算法结合形成一种新的信息融合方法,综合考虑地震、测井、录井和钻井各参数与地层可钻性级值的关系,将提出的信息融合方法结合多维异构空间模型应用于地层可钻性级值预测中。通过机械钻速方程,利用遗传算法将与钻速有关的钻进特征参数提取出来。根据实时钻进特征参数的变化趋势,提出了一种基于钻进特征参数的异常地层压力识别方法。通过分析钻进特征参数中压差系数与地层压力的关系,提出了一种基于钻速方程的地层压力随钻监测方法。通过分析地震层速度与地层压力的关系,提出了一种改进的Fillippone地层压力预测模型。以地区地质信息及已钻井数据为依据进行模型参数初始化,融合随钻获取的有关地层压力数据及钻井液密度信息利用贝叶斯推理方法对预测模型参数进行实时更新,从而提高了钻头前方地层压力预测的准确性,同时降低了预测的不确定性。为了更精确地解决测井岩性识别问题,提出一种改进的Adaboost-SVM方法。首先,利用核主成分分析方法对实际测井数据进行特征提取,然后利用提取的特征和岩性剖面资料训练支持向量机,最后将若干个支持向量机弱分类器通过Adaboost算法进行集成,建立测井岩性识别的Adaboost-SVM模型,并将该岩性识别模型与基于QPSO-SVM的岩性识别模型的岩性识别结果进行了比较。在分析了单独利用沃尔什变换或高斯模型进行测井曲线自动分层优缺点的基础上,提出了一种基于沃尔什变换和高斯模型的多测井曲线融合分层方法。该方法融合了沃尔什变换简单、快速的特点,同时又利用了高斯模型对沃尔什变换分层的边界进行校正,实现了测井曲线由粗到细的分层过程。实例应用表明,该方法能够利用多条测井曲线进行自动分层,效果良好,验证了模型的可靠性和实用性。将支持向量机和变异函数结合,提出了一种新的空间插值方法,可用于钻井地质特征参数的插值估计。该方法将变异函数作为支持向量机目标函数的约束条件,不但利用了变异函数的空间相关结构重建能力,而且还保留了支持向量机较强的非线性回归能力,同时考虑了空间变量的属性相关性和空间相关性,取得了较好的插值重构效果。更多还原

【Abstract】 A great quantity of complex and uncertain factors exist in petroleum exploration and exploitation, especially in the drilling engineering. The different geological conditions, such as the change of lithology, formation drillability and pore pressure, may be encountered at any time during drilling, which can bring complex factors to the drilling process. The inaccurate understanding of the complex drilling environment underground will affect the drilling construction, and even cause drilling accidents and endanger the personal safety. In this paper, first the multi-source information in petroleum exploration and exploitation is analyzed, and then the estimation and prediction of drilling geologic characteristic parameters, such as the formation drillability, pore pressure, formation lithology etc, are emphatically investigated in detail. The main contributions are as follows.The information source and expression pattern in petroleum exploration and exploitation are studied, and then the relation between the multi-source information and the drilling geologic characteristic parameters is analyzed. After that, the multi-dimensional heterogeneous spatial model is established and the multi-source information is characterized with a knowledge expression system.The relation between mud logging, well log data and formation drillability is analyzed, and a novel method for predicting formation drillability based on relevance vector machine (RVM) is proposed. Then the prediction model for formation drillability is established by training the RVM. The quantum particle swarm optimization algorithm (QPSO) and support vector machine (SVM) are combined to form a new information fusion method. Then this method with the multi-dimensional heterogeneous space model is applied to formation drillability prediction.The drilling characteristic parameters related to the drilling penetration rate are extracted with genetic algorithm using the drilling rate model. After that, an abnormal formation pressure detection method based on drilling characteristic parameters is proposed by analyzing the trend of these real-time drilling characteristic parameters. Then a pore pressure monitoring while drilling method is proposed by the analysis of the relation between differential pressure coefficient and the pore pressure according to the extracted drilling characteristic parameters. A modified Fillippone pore pressure prediction model is proposed through the analysis of the relation between seismic interval velocity and pore pressure. Firstly, the parameter in the model is initialized according to the region geological information and the data of drilled well. Then the acquired pore pressure data and the mud weight information while drilling are incorporated into the model as the constraint, which makes it possible to update the parameters in the model with Bayesian inference method in real time. Thereby, the pore pressure prediction accuracy is increased and the uncertainty of the prediction is reduced at the same time.To address the well log lithology identification issue accurately, a novel modified Adaboost combined with SVM method is proposed. First, the kernel principal component analysis algorithm is utilized to extract the feature of the actual well log data. Then the support vector machine is trained with the extracted feature and lithologic profile data. After that, the lithology identification model is established based on Adaboost-SVM by boosting a number of weak support vector machine classifiers with Adaboost algorithm. Finally, the lithology identification results with Adaboost-SVM algorithm are compared with the results using QPSO-SVM algorithm.By analyzing the advantages and disadvantages of each method based on Walsh transform or Gaussian model for automatic segmentation of well logs respectively, a method for joint application of the Walsh transform and Gaussian model for automatic segmentation of multiple well logs is proposed. On one hand, the method maintains the feature of simplicity and rapidity of Walsh transform. On the other hand, it adjusts the segmentation boundary with Walsh transform by Gaussian model. Thereby, a coarse-to-fine segmentation of well logs can be achieved. The experimental results show that the proposed method can better perform the segmentation with multiple well logs and has higher practicability and reliability.Based on the combination of support vector machine and the semivariogram, a spatial interpolation method is proposed in the paper, which can be used to the drilling characteristic parameters interpolation estimation. Considering the semivariogram as the constraint of the objective function of support vector machine, this method can not only make use of the spatial correlation structure reconstruction ability of semivariogram, but also retain the strong nonlinear regression ability of support vector machine. The attribute correlation and spatial correlation of the spatial variation are taken into consideration simultaneously, which results in better interpolation results.更多还原