【作者】 贺懿；

【导师】 刘怀山；

【作者基本信息】 中国海洋大学 ， 海洋地球物理学， 2008， 博士

【摘要】 目前,应用地震勘探进行岩性油气藏的预测仍处于探索及技术的积累阶段,多数技术尚处于理论模型的研究探索阶段,在生产实践中缺乏普遍性和针对性。而地震反演技术一直是地震勘探中的一项核心技术,其目的是用地震反射资料,提取地震属性,反推地下的波阻抗或速度的分布,估算储层参数,并进行储层预测和油藏描述,为油气勘探提供可靠的基础资料。利用地震属性能从地震数据中获得其他方法无法提供的信息,预测储层厚度、储层孔隙度、渗透率、饱和度等参数的空间变化。但是,这些参数只能通过钻井测定,而无法直接由地震数据确定。这就需要建立测井数据获得的各种参数与地震数据之间的关系,这些储层参数与地震信息之间的关系并不具有一一对应的特性,很难用一种精确的算法来准确地描述它。目前,常用的一种方法就是利用地震属性提取技术将不同的地震属性从地震数据中提取出来,然后在主要的地震属性与测井得到的储层参数之间建立起一种对应关系,由点及面,通过地震数据将这种关系展布到三维空间,以此预测储层参数。本文在对地震储层预测技术的国内外研究现状、人工神经网络的研究与发展现状、模拟退火算法以及粒子群算法的研究现状进行了充分调研的基础上,针对储层预测研究内容,以技术创新、方法可行和效果优先为原则,与实际相结合,系统地研究和分析了地震属性提取与优化方法、约束波阻抗反演技术、声波重构技术等内容,提出了多曲线声波重构方法、基于免疫规划的模拟退火神经网络技术以及竞争粒子群-神经网络技术,并将这些方法技术应用于实际资料。研究了地震波阻抗反演方法原理并分析比较了不同反演方法,以选择适合实际情况的波阻抗反演方法。详细叙述了反演前数据的准备工作,包括数据的高分辨处理、高精度的构造解释以及测井数据的处理。对所选稀疏脉冲反演方法的原理、流程作了深入的学习和分析,并针对已有声波重构方法存在的局限性,提出了同时利用多种对储层敏感的非声波测井信息,对原始声波测井曲线进行重构的技术。由于该重构技术利用了不同测井信息,保留声波曲线所有频率信息,并将参与重构的曲线所有信息融合到声波曲线中,既保持声波曲线原有的时深关系不变,又能实现显著提高储层与围岩之间的差异,突出储层特征,反映地层岩性的细微差别,因此,该技术可以较好地解决砂泥岩薄互层储层定量预测中的岩性识别、薄层分辨等难题。将多曲线声波重构技术用于稀疏脉冲反演进行地震储层反演,获得较好的波阻抗剖面,用于储层的追踪和识别,能有效地预测未知储层的空间展布情况。在比较分析地震数据属性分类、提取及优化方法的基础上,选用合理科学的地震属性提取方法和优选方法,对地震数据进行属性的提取和优化,为后续的储层预测提供更符合实际地质情况的属性数据。研究分析了传统前馈反向传播神经网络原理和模拟退火算法,了解掌握了模拟退火算法存在着搜索和收敛速度较慢,同时在每一个温度下,其解的搜索空间被限定在较小的范围内,从而导致全局最小值不能真正地被搜索到的缺点。由于免疫规划算法是基于进化规划的一种全局搜索算法,将其和模拟退火算法结合,能够克服模拟退火搜索收敛速度慢的缺点,并获得一种有效地搜索全局最优解的搜索算法。针对反向传播算法的不足和模拟退火算法的缺点,研究提出了基于免疫规划的模拟退火算法,并提出了改进模拟退火算法替代BP算法训练网络参数的神经网络技术,同时分析了该网络的学习算法,最后利用该方法对提取优选的地震属性进行储层参数预测应用,获得了较好的效果。研究分析了粒子群算法的原理及其在进化后期存在速度变慢以及过早收敛(即早熟)的缺点,在此基础上提出将进化规划中子代与父代群之间的竞争思想引入粒子群的搜索和收敛过程中,以此提高粒子群算法的搜索速度,并防止过早收敛,构成了竞争粒子群算法。应用竞争粒子群算法替换BP算法用于前向神经网络的参数训练,以便克服反向传播神经网络由BP算法带来的缺点,同时利用该算法进行网络结构的优化训练,使网络的预测性能有较大提高,形成一种适合于储层参数的粒子群-反向传播神经网络预测方法,在实际应用中获得了比较满意的结果。本文研究成果对石油的勘探、开发和科研具有重要的指导意义与实用价值。更多还原

【Abstract】 At present,the prediction of lithologic hydrocarbon reservoir using seismic exploration still is in accumulation period of exploratory and technology,and most of those technologies is in the exploratory investigation of theoretical model,too. However,seismic inversion technique has always been one of core techniques of seismic exploration,in which seismic reflection data has been used in extracting seismic attributes and derivating the distribution of wave impedance or velocity that have been used to estimate the reservoir parameters,realize reservoir prediction and description and apply the reliable basic data for hydrocarbon explorationUsing seismic attributes can extract the information that others methods have never obtained,which has been applied to predict the three-dimensional distribution of reservoir thickness,porosity,permeability,saturation and so on. However,these parameters can not be measured directly but well logging,which requires to build the relationship between various parameters obtained by logging and seismic data. The relationship of these parameters and seismic data is not one-to-one correspondence,it is hard to describe the relationship using an accurate algorithm precisely. Presently,there is one of methods in common use that extracts various seismic attributes from the seismic data using extracting technique,builds a relation between primary attributes and parameters obtained by logging,distributes the relation over the whole three-dimensional space by seismic data from point to surface,and finally predicts reservoir parameters.Based on investigating sufficiently to the present researching situation of reservoir prediction,Artificial Neural Network,Simulating Anneal Algorithm and Particle Swarm Optimization in home and abroad,the paper arms at the research contents of reservoir prediction,demonstrates and analyzes seismic attributes extraction and optimization method,constrained wave impedance inversion technique and acoustic wave curve reconstruction technique etc,develops multi-parameters rebuilding acoustic wave curve method,SA-ANN technique based on Immune Programming,and Competition PSO-ANN technique,moreover,applies them to actual data.Armed at the limit in reconstruction methods existed,this paper proposes one method that uses multi-non-acoustic logging information that is sensitive to reservoir,and rebuilds the original acoustic wave curve. Because the method uses different logging information,saves all of the frequency information of acoustic wave curve,syncretizes all of the information of the curves participated in reconstruction into curve rebuilt which not only keeps the original time-depth relation of acoustic wave curve unchanging,but also realizes enhancing the difference between reservoir and surrounding rock,giving prominence to the property of reservoir and imaging the hairline of formation lithology,this technique can resolve the difficult problem about lithologic discrimination and thin bed resolution in reservoir quantitative forecast of the thin interbeds of sandstone and mudstone.After completely comparative analyzing the classification , extracting and optimization methods of seismic data attributes,the reasonable and scientific extracting and optimization methods of seismic data attributes are choose and used in extracting and optimization of attributes,offer reservoir prediction the attributes that are much more in accord with actual geology situation.Deeply studying and analyzing theory of Back Propagation Neural Network and SAA,the shortage that the velocity of searching and constringency is slow,at the same time,the searching space of the answer of SAA is limited in little area under per temperature so that it is likely to induce the global optimization answer not to be searched is known. Because Immune Programming is one of global searching algorithm based on Evolutionary Programming,the new algorithm that can effectively search global optimization answer will be gained and overcome the shortage of SAA if Immune Programming is combined with SAA. For conquering the flaw of BPNN and SAA,the paper researches and proposes a SAA based on Immune Programming,moreover,brings forward a NN technique in which network parameters are trained by replacing BP algorithm with SAA improved. At the same time,the study algorithm of this NN is researched and analyzed. The satisfactory effect is acquired when it is used in reservoir parameters prediction for seismic attributes optimized.The limitation of PSO is brought forth enough after the principle of PSO is studied. There are two primary limitations that the searching velocity will be slow down in the evening of evolutionary and PSO trend to over-constringency (prematurity) so that only the local optimization answer will be obtained. So the competition idea between filial generation and father generation in the Evolutionary Programming Algorithm is introduced into the searching and constringency course of PSO to form a competition PSO in order to enhance the velocity of searching and avoid the prematurity phenomenon. When appling competition PSO to train the network parameters can overcome the inherent shortcoming existed in BPNN,using this optimization algorithm optimizes the structure of network in order to improve the capability of network. The network trained is applied to actual data to predict reservoir parameters,the predicting result accords with actual logging data basically.The result and production of study in this paper have direction significance and applied value for oil exploration,exploitation and scientific research.更多还原