基于图形处理的多点地质统计算法及模型评价

发布时间：2017-12-31 18:29

本文关键词：基于图形处理的多点地质统计算法及模型评价　出处：《中国科学技术大学》2016年博士论文　论文类型：学位论文

【摘要】：储层建模是进行储层相关研究最基本的一环。建立精确的三维储层数字化模型是进行储层特征描述、储层评价、储层预测和相关应用的基础。随着勘探技术与数据采集能力的飞速提升和储层数值模拟在实际应用中重要性的逐步加强,现代储层研究中对建立的储层模型的规模和精度要求在不断加强。而多点地质统计学算法作为储层建模的主要工具之一,往往存在着计算耗时长、内存需求大、参数敏感性强、模型评估困难等限制,建立准确反应储层特征的高质量、大规模储层模型还存在着一定困难。本文针对这些缺陷,从多点地质统计学和计算机图形领域的相似点和交叉点出发,将计算机图形处理中的相关方法引入多点地质统计学,分别提出了基于计算机图形处理器通用计算能力的高阶累积量表征的并行计算方法,此并行算法在保证计算精度的同时大大提高了计算效率；提出了基于计算机图形学图形切割算法的新型高效模式模拟算法,此算法克服了参数依赖改善了模型效果；提出了扩展终端定义方法将基于图形切割的多点算法推广到点条件数据模拟,打破了图形切割算法不能用于点条件数据约束的限制,克服了模式模拟算法中点数据条件模拟的困难,实现了便捷的保证局部条件数据一致性的条件模拟算法。此外,本文还展开了多点地质统计学储层模型的地质模式重构能力、变化性和水利连通能力等方面的研究,提出了融合指数概念用于模式模拟模型的量化评估标准。本文的主要成果及创新如下：1.针对多点地质统计学空间高阶累积量的计算量巨大的缺点,提出了基于计算机图形处理器(GPU)通用计算能力的并行算法：i) 提出了计算统一构架(CUDA)框架下的GPU并行算法。通过将空间累积量的计算划分为每个步长下的空间矩单元的计算和将空间矩转化为空间累积量的两部分,提出了两阶段并行策略,显著提高了计算效率。ii) 针对累积量计算中需要大量内存,提出了CUDA内存优化策略。充分考察了GPU中的各种内存间数据交换和通信能力,减少了GPU和CPU之间的通信瓶颈的数据传输,加快了GPU内存的利用率,进一步提高了并行计算效率。iii) 改进并采用了基于GPU并行的递归算法,克服了GPU原子计算造成的计算能力降级。此外还进行了线程分配对算法效率的敏感度分析。iv) 考察了GPU编译器的浮点数误差,通过修改运算顺序和变量类型,提高厂GPU并行计算精度。2.提出了基于计算机图形切割算法的条件模式模拟迭代算法：i) 将计算机图形切割算法术引入多点地质统计学算法。用最小剪切方法缓解模式模拟中的边界问题。ii) 提出将图形切割算法中的剪切值作为一个模拟效果参考量,用于模型的连续迭代修正。iii) 将条件模式模拟中的条件数据作为模拟参考量,使用迭代方法达到精确的点数据吻合。同时提出了扩展终端的概念和定义方法,确保了点条件数据模拟的局部一致性。3. 开展了对基于图形切割的多点地质统计模式模拟算法在建模效果和性能上的研究：i) 提出了融合指数的概念,作为对训练图像整体复制量的一种量化标准,用于度量模式模拟算法生成模型的变化性。ii) 对所提模式模拟中的参数进行了敏感性分析,证明所提方法的收敛性,即对参数不敏感的优势。iii) 采用不同的指标,包括目标体距离函数、空间统计平均值和统计残差以及水利连通性流动响应等多个方面,将所提算法与现有的逐点模拟算法以及模式模拟算法进行了全方位的对比,得到所提算法能够产生至少与逐点模拟方法相同的变化性以及更好的连通性的优点。同时,性能对比显示,所提算法能大幅度减少计算时间。
[Abstract]:Reservoir modeling is the reservoir related to the most basic research part. To establish a precise 3D reservoir model of digital reservoir description, reservoir evaluation, reservoir prediction and correlation based applications. With the rapid development of exploration technology and data acquisition capability and reservoir numerical simulation value in practical application the gradual strengthening of modern reservoir study on scale and accuracy of reservoir model requirements are continuously strengthened. And the multiple point geostatistics algorithm as a major tool of reservoir modeling, there are often time-consuming calculation, memory requirements, parameter sensitivity, evaluation model is difficult to establish accurate, high quality the reaction characteristics of the reservoir, the large-scale reservoir model still exist some difficulties. In view of these defects, starting from multiple point geostatistics and computer graphics similarities and intersection calculation The related method for graphics processing in the introduction of multiple point geostatistics, put forward calculation methods of computing capability based on general computer graphics processor of high order cumulant characterization of parallel, the parallel algorithm in improving computing accuracy and greatly improves the computational efficiency; simulation algorithm is presented for efficient mode of computer graphics cutting algorithm based on this algorithm overcomes the parameter dependent model is proposed to improve the effect; extended terminal definition method of multi point algorithm is extended to point condition data graphics based on cutting simulation, breaking the graph cut algorithm cannot be used to condition data constraints, the model simulation algorithm to overcome the difficult conditions of point data, realizes the data to ensure the local conditions the consistency of the simulation is convenient. In addition, this paper also carried out multiple point geostatistics in reservoir model The geological model reconstruction ability, and changes of water conservancy connectivity and other aspects of the fusion index concept for quantitative evaluation standard model simulation model is presented. The main achievements and innovations of this paper are as follows: 1. for the multiple point geostatistics spatial calculation of high order cumulant amount huge disadvantages, proposed the computer graphics processor (based on GPU) parallel algorithm for computing capacity of: I) proposed a unified framework (CUDA) under the framework of GPU parallel algorithm. By calculating the space dividing cumulant is a rectangular unit of each step under the calculation and two part of the spatial moment into space accumulation, put forward two stage parallel strategy. Significantly improve the computational efficiency of.Ii) for a large amount of memory need cumulant calculation, proposed CUDA memory optimization strategy. Fully examine the ability of various memory in GPU data exchange and communication, To reduce the data transmission bottleneck between GPU and CPU, to speed up the use of GPU memory rate, to further improve the efficiency of parallel computing and improved.Iii) using a recursive algorithm based on parallel GPU, overcome the downgrade computation ability of GPU atom computing caused. In addition to the analysis of sensitivity of.Iv thread allocation efficiency of the algorithm the effects of floating-point error) GPU compiler, by modifying the operation sequence and variable type, improve the accuracy of the.2. plant GPU parallel iterative algorithm is proposed to simulate the condition of model of computer graphics cutting algorithm based on computer graphics: I) cutting algorithm into multiple point geostatistics algorithm. Using the minimum shear method to alleviate the.Ii boundary problem in the simulation of the graph cut algorithm) in the shear value as a reference for the simulation results, continuous iterative correction.Iii model) will die The simulation conditions in the simulation data as the reference, using iterative method to point accurate data well. And puts forward the concept and definition of the terminal expansion method, to ensure that the local consistency conditions.3. data simulation is carried out on the study of geological statistics model to simulate the algorithm in the modeling results and the performance of the graph cut based on: I) put forward the concept of fusion index, as a quantitative criterion for training image reproduction in whole amount, for change of.Ii metric model generation algorithm model of the proposed simulation model) parameters in the sensitivity analysis, prove the convergence of the proposed method, which is not sensitive to the parameters the advantages of.Iii) using different indicators, including the object distance function, the average value of spatial statistics and statistics as well as water flow connectivity and other aspects of the response, the proposed algorithm is compared with the existing The algorithm of point by point simulation and pattern simulation algorithm are compared in an all-round way. It is concluded that the proposed algorithm can produce at least the same variability and better connectivity as point by point simulation method. Meanwhile, the performance comparison shows that the proposed algorithm can greatly reduce the computation time.

【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：O212.1;TP391.41

【相似文献】