【作者】 高美娟；

【导师】 翟云芳；

【作者基本信息】 大庆石油学院 ， 油气田开发工程， 2005， 博士

【摘要】 石油的产生有着不可重复性,不可实验性的特点。人们对地下油气储层的认识存在相当程度的模糊性和不确定性。由于某些不正确的认识会带来巨额生产资金消耗。研究准确识别与预测储层及参数空间变化的方法,在油气勘探开发中都具有重要的实际意义,可以提高生产效益,为国家节约巨额的资金。本文以油田的科技攻关项目为背景,以准确、清晰地再现地下油气储层及参数空间展布,提高油气勘探开发的准确率,节约生产成本,提高生产效率为目标,将计算机模式识别技术和现代预测理论与石油生产相结合,围绕对地下油气储层及参数的识别与预测进行一系列的研究工作。主要内容如下:1.在系统分析研究人工神经网络原理的基础上,对误差回传神经网络(BP)的结构、计算公式进行了深入研究。针对地下储层参数与地震特征参数是复杂的非线性关系,很难用精确的表达式表示出来这一问题,指出了神经网络以其自身的非线性逼近能力是解决上述问题的有利工具,同时给出了神经网络进行储层参数预测的过程描述。2.在深入分析径向基函数(RBF)神经网络和遗传算法理论的基础上,提出了一个基于自适应遗传算法的径向基函数神经网络的结构优化算法,并将该算法成功地应用于储层参数的横向预测,取得了非常好的效果。该算法将基函数的中心参数和宽度参数编成染色体,将网络结构优化和参数学习分两个阶段进行,训练和进化。先用梯度下降法学习某一染色体对应的网络的中心和宽度参数,而后用最小二乘法学习网络的线性权值和偏移常数;再用遗传进化算法优化隐节点数。通过这两个过程的交替进行,得到满足误差要求的具有最小隐节点数的并且隐节点基函数具有不同宽度参数的RBF神经网络。3.深入研究分析小波变换理论和边缘检测原理,并将地震数据剖面看成是二维数字图像,地震剖面上的反射界面相当于二维数字图像的边缘。首次提出将小波变换模极大边缘检测技术应用到地震剖面图像的特殊处理上,使用于储层研究的地震资料具有较高的质量,即高信噪比、高分辨率。从而使计算机能更准确、可靠地预测储层参数的空间展布,提高勘探开发的精确度。用该方法对实际资料进行了处理,取得了较好的效果。4.在深入研究分析小波神经网络的结构、性质、算法和逼近能力的基础上,同时通过小波神经网络与常用网络的比较,提出了用于储层参数预测的小波神经网络优化算法,得到了较为优化的网络结构,并给出了系统的学习算法。由于人工神经网络具有自学习、自适应、鲁棒性、容错性和扩充性能,而小波分析具有时频局部特性和变焦特性,将两者的优势相结合,使得小波神经网络具有较强的逼近、容错能力。实际资料预测表明,本文提出的用于储层参数预测的小波神经网络具有收敛速度快,逼近非线性能力强的优点。5.在深入分析模糊系统理论和神经网络的基础上,针对神经网络和模糊系统各自的优点和不足,对模糊系统和神经网络的融合技术进行了深入研究,结合火山岩储层识别预测的特点和难点,提出了用于火山岩储层识别预测的模糊神经网络系统,同时给出了系统的结构和参数学习算法及识别预测火山岩储层的步骤。在方法研究过程中,充分利用现有的样本数据,发挥神经网络的自适应学习能力和模糊系更多还原

【Abstract】 Petroleum cannot be reformed by experiment, and man cannot made it repeatedly underground. Man havesome unclear and indeterminacy cognition about underground oil reservoir, it make waste a lot of money everyyear. So it is very important that man realize the structure and parameter distribution of oil reservoir clearly. Itcan heighten production efficiency and save a lot of money .In this paper , some scientific research item of oil field are used as the background. And it is used asobject that the modality of oil reservoir is reappeared clearly and the accuracy rating of exploration isimproved. We study the pattern recognition theory and the methods for accurate fast prediction the parametersof oil reservoir. The main contents are as follows:1. After analyse and study systematically artificial neural network theory, the structure and computeformula of BP is studied. Considering the nonlinear relation between reservoir parameter and seismiccharacteristic parameters, the neural network is a advantageous tool to solve above problem, the coursedescription of reservoir parameter predicting using neural network is provided.2. Based on studying radical basic function neural network and genetic algorithm, a optimize method ofradical basic function neural network based on self-adapting genetic algorithm is proposed, and this method isapplied successful to the reservoir parameter predicting and obtain better effect. In this method, the centerparameter and width parameter are considered chromosome, the network structure optimize and parameterlearning are divided into training and evolution two parts. At first, learning the center parameter and widthparameter using gradient descend algorithm, then learning the linear weight value and offset constant usingleast square algorithm. Secondly, optimize the hidden node numbers using self-adapting genetic algorithm.Finally, we can obtain the RBF neural network which satisfy the error demand by alternate executing abovetwo courses.3. The wavelet transform theory and edge detection principle is in-deep studied, the seismic sections usedfor seismic interpretation are actually regarded as some 2D images, so the reflection interface of seismicsection can be considered as edges on the seismic section images. The wavelet transform Modulus-Max edgedetection algorithm which applied to seismic section processing is proposed first, its purpose is to eliminatethe random interference and enhanced the quality of seismic section and the signal-to-noise rate and resolutionof seismic signal, furthermore, predicting the parameter distribution of oil reservoir accurately and credibility.This algorithm is used to process actual data and obtain better effect.4. Based on analyse and studying the structure and property and approach ability of wavelet neuralnetwork , a method of wavelet neural network which is used to predict reservoir parameter is proposed ,meanwhile, the optimize network structure and learning algorithm of network are presented. Artificial neuralnetwork has the capability of self-learning and self-adapting and approaching nonlinear, the wavelet transformhas good partial property both in time domain and frequency domain, so the wavelet neural network whichcombining the advantage of ANN and wavelet has stronger approaching nonlinear capability and allowanceerror capability. The actual data predicting shows that this method provides higher predict accuracy and fasterconvergence speed.5. The fuzzy system theory is studied, considering to the advantage and disadvantage of the neuralnetwork and fuzzy system, the fusion problem of fuzzy system and neural network is in-deep studied,combining the salient feature and difficulty of volcanic reservoir, a fuzzy neural network system which appliedrecognition and predicting the volcanic reservoir is presented, meanwhile, the system structure and parameterlearning algorithm and predicting steps are given. In this thesis, stronger self-learning ability of neural networkand good express ability of fuzzy system are combined to recognise and predict the volcanic reservoir. Actualapplication result appear that this method recognise and predict the volcanic reservoir accurately and fast.6. Based on studying the statistical learning theory and machine learning and support vector machineprincipal, considering the ANN has some problems in application such as converge to local minimum, theoverfitting and the structure of ANN is always decided by experience because it doesn’t have a good guidingtheory, especially, when the number of the training sample is not enough, the predicting accuracy will beinfluenced, a method of recognition and predicting the volcanic reservoir based on regression support vectormachine(SVM) is proposed first. SVM is a newly developed technique which based on statistical learningtheory, it adopts Structure Risk Minimization principle which avoids the disadvantages of ANN. In this thesis,the recognition and predicting of regression algorithm based on SVM is studied, the fast learning algorithm isstudied and the concrete realization is presented. Actual application result shows that this method remedy thescarcity of few sample data, overcome the disadvantages of traditional neural network, therefore, theconverging speed and the predicting accuracy of system are enhanced.7. Programming some programmes using C++ such as the radical basic function neural network based onself-adapting genetic algorithm, the wavelet transform Modulus-Max edge detection algorithm which appliedto seismic section processing, the wavelet neural network algorithm which is used to predict reservoirparameter, the fuzzy neural network system which applied recognition and predicting the volcanic reservoir,the recognition and predicting the volcanic reservoir based on regression support vector machine algorithm.更多还原