【作者】 徐慧；

【导师】 杨永国；

【作者基本信息】 中国矿业大学 ， 地球信息科学， 2013， 博士

【摘要】 煤层气产能是衡量煤层气井潜在产气能力的综合指标，产能的高低直接影响煤层气项目的经济效益。因此建立有效的煤层气产能预测模型，对煤层气井的勘探开发有着重要的指导意义。煤层气赋存于煤储层中，其产出过程由多个地质因素决定且各因素之间关系复杂，难于建立精确的数学表达式来描述其动态的生产过程。因此本论文采用目前广泛应用于预测控制等领域的支持向量回归机以及改进的粒子群优化算法来建立地质因素与产能之间的非线性函数映射关系，以实现对煤层气井产能进行预测及控制的目的。支持向量回归机模型的建立不仅需要一定数量的样本数据进行训练和测试，同时为了建立高质量的预测模型，需要对模型中的参数设定最优的取值，因此选用粒子群优化算法对参数进行优化。粒子群优化算法目前已经广泛地应用于各个领域，但是由于它自身的进化特点导致其在寻优过程中容易陷入局部收敛。为了解决该算法易于陷入局部收敛的问题，本文主要提出了三个改进的粒子群优化算法。（1）基于子维进化的粒子群优化算法从标准粒子群优化算法的进化策略入手，将种群中粒子的进化策略从粒子的整体进化改变为粒子的每一维依次进化。同时当种群陷入局部收敛时，采取对多样性较差的子维进行重新初始化的操作。无论是对简单的单峰函数还是复杂的多峰函数进行优化，相较于标准粒子群优化算法该算法均具有较好的寻优性能。（2）基于免疫机制的混合粒子群优化算法融合了人工免疫算法和基于子维进化的粒子群优化算法，将进化过程分成两个阶段，第一阶段采取人工免疫优化算法进行全局寻优，为下一阶段的寻优提供质量较高的初始种群。第二阶段采取基于子维进化的粒子群优化算法在质量较高的初始种群的基础上进行进化寻优，因此该算法具有更高的寻优效率。（3）多种群协同进化的粒子群优化算法在Agent的协同作用下，分别由人工免疫算法、混沌算法，子维进化的粒子群优化算法同时进化，在粒子群算法陷入局部收敛时，共享其他两个算法的最优值，以较高的质量跳出局部收敛，为进一步的寻优工作打下良好的基础，同样该算法也具有更高的寻优效率。通过标准数据集Boston Housing作为数据样本，将两种改进的混合粒子群算法应用于优化支持向量回归机模型中的参数，结果表明多种群协同进化的粒子群优化算法更适用于优化模型参数。通过选定沁水盆地南部樊庄区块的20口煤层气垂直井的相关数据，利用改进的粒子群优化算法优化支持向量回归机，建立煤层气产能预测模型，并与BP神经网络以及支持向量回归机的预测结果进行比较，结果表明改进的粒子群优化算法优化支持向量回归机建立的模型具有更高的预测精度。同时根据20组样本数据对参与建立模型的5个地质因素分别进行了实验，分析了它们对产能的影响。更多还原

【Abstract】 CBMproductivity is the comprehensive indicatorfor measuring potential gasproduction of CBM wells, while the productivity directly affects the economicbenefits of CBM project. Therefore, development of effective CBM productivityprediction model has important guiding significance to the exploration anddevelopment of CBM Wells.CBM is saved in the coal reservoir, and its productivity is determined by manygeological factors, and the relationship is complex, so it is difficult to establishaccurate mathematical expressions to describe the dynamic process. This paper adoptssupport vector regression machine which is widely used in predictive control andother areas currently and the improved Particle Swarm Optimization Algorithm tocreate the nonlinear function mapping relationship between geological factors andproductivity, so as to realize prediction and control of CBM well productivity.The development of support vector regression machine model requires a certainamount of sample data for the training and test of structure. In order to develop thehigh quality prediction model, it is required to set the optimum value of theparameters in the model. Therefore, in order to optimize the parameters in the model,particle swarm optimization algorithm is selected for the optimization of parameters.Although particle swarm optimization algorithm is widely used in various fields,its evolution characteristics easily lead to local convergence. In order to solve theproblem that the algorithm easily falls into local convergence, this paper basicallyproposes three improved particle swarm optimization algorithms.Particle swarm optimization based on the evolution of sub-dimensions startsfrom the evolutionary strategy of standard particle swarm optimization algorithm, andit changes the evolutionary strategy of particles of the population from overallevolution of the particles to each dimension of the particles for successive evolution.When the particle is trapped in local convergence, the sub-dimensions with poordiversity valueis are reinitialized. Regardless of whether they are used in simpleunimodal function or complex multimodal function optimization, compared withStandard Particle Swarm Optimization Algorithm, this algorithm has betteroptimization performance.Hybrid Particle Swarm Optimization based on immune mechanism integratesArtificial Immune Algorithm and Particle Swarm Optimization Algorithm which is based on sub-dimensional evolution, the evolution process is divided into twostages.The first stage is to use Artificial Immune Optimization Algorithm for globaloptimization, so as to provide high quality initial population for the next phaseoptimization. The second stage is to use Particle Swarm Optimization Algorithmwhich is based on sub-dimensional evolution for evolutionary optimization based onmultiple high quality initial population, Therefore, it has higher efficiency ofoptimization.Hybrid particle swarm optimization algorithm of multiple-populationCooperating evolution can be evolved at the same time by Artificial ImmuneAlgorithm, Chaos Algorithm and Particle Swarm Optimization Algorithm of sub-imensional evolution. Agent records current global optimal value obtained from thesethree algorithms. When Particle Swarm Optimization Algorithm gets into localconvergence, it jumps out of local convergence with high quality through the recordedcurrent global optimal value, so as to lay a good foundation for further optimizationwork. Meanwhile, the algorithm also has higher efficiency of optimization.The standard data set Boston Housing is selected as data samples. Two improvedhybrid particle swarm algorithms are applied to optimizing support vector regressionmachine parameters.The results show that the Particle Swarm Optimization Algorithmof multiple-population synergy is more applicable for optimizing model parameters.Through selecting20groups of related data for the20CBM vertical wells ofFanzhuang block to the south of Qinshui basin o develop CBM productivityprediction model with the Particle Swarm Optimization Algorithm Optimizationoptimized by using the support vector machine regression model. Compares with theprediction result of BP neural network and support vector regression machine, theresults show that the improved Particle Swarm Optimization Algorithm Optimizationand support vector machine regression model has higher precision of prediction.Meanwhile, five geological factors which are involved in model development aretested respectively according to the20groups of sample data, the impact of fivegeological factors on CBM productivity are analyzed.更多还原