【作者】 吴秋波；

【导师】 王允诚；

【作者基本信息】 成都理工大学 ， 油气田开发工程， 2011， 博士

【摘要】 储层预测是建立精确油气藏地质模型，准确估算油气储量，确定合理开发方案的基础工作，不仅可用于油气勘探，而且对于指导油气藏特别是复杂隐蔽油气藏或岩性油气藏的开发具有重要意义，已成为国内外学术界和工业界共同关注的关键科学问题。通过井间对比或插值外推对储层横向展布或储层参数予以精确描述是比较困难的。联合测井和地震信息的反演技术和地震属性技术作为重要的储层预测方法，已延伸至油气开发阶段，成为油气藏描述技术的重要组成部分。反演的实质是求取非线性目标函数的极值问题，常规的线性反演有可能陷入局部极值，或依赖于初始模型的选择，从而影响反演的可靠程度。智能计算方法如遗传算法或模拟退火等为求解储层参数反演这样的非线性问题提供了新的思路和手段，但也存在不足，比如求解速度慢，或对于复杂反演问题，也可能会陷入局部极值等。智能计算方法中的神经网络技术常用来建立地震属性和储层参数间非线性映射关系，进而预测储层参数。误差反向传播网络学习算法实际上是梯度算法，收敛速度慢，也可能会陷入局部极值。针对上述问题，论文以改善储层非线性反演的求解效率和寻优性能，寻求高效的学习算法以提高储层参数预测效果为目的，研究基于粒子群优化算法的储层参数非线性反演方法和基于混合学习法神经网络的多属性储层参数预测技术，并应用这些方法预测川东DCH构造飞仙关组鲕滩储层的空间展布规律和孔隙度等参数，取得了如下成果和认识：1、深入地研究了粒子群优化算法原理、收敛性条件、参数设置原则等基本理论问题，针对优化问题求解，粒子群优化算法对解的更新更具目的性，收敛速度快。提出的混沌惯性权重调整策略的粒子群优化算法，进一步提高了算法的收敛速度；将多父体交叉算子和群体爬山思想融入粒子群优化算法中，很好地提高了粒子群优化算法的寻优性能，上述改进为解决储层参数反演问题提供了新的优化工具。在此基础上，研究开发了基于粒子群优化算法及其改进算法的储层参数非线性反演方法，模型试算和实际资料处理结果表明新方法较传统非线性方法相比，在求解效率上有了较大的提高，效果较好。2、深入地研究了误差反向传播网络和径向基函数网络及其学习算法。研究了粒子群优化算法训练误差反向传播网络和径向基函数网络。提出了融合梯度法和粒子群优化算法的径向基函数网络混合学习算法，在求解精度和效率方面有了较大的提升。在此基础上开发了基于混合学习法前向神经网络多属性储层参数预测方法，应用于储层孔隙度参数的横向预测，在收敛速度上有了较大的提升。3、应用基于粒子群优化算法的储层参数非线性反演方法反演了川东DCH构造飞仙关组鲕滩储层速度，应用混合学习法神经网络多属性储层参数预测方法预测了该储层孔隙度。在此基础上制作了鲕滩储层厚度、孔隙度、储能系数预测图。综合地质、地震和测井信息研究，认为沉积相带控制了DCH构造飞仙关组鲕滩储层发育程度和范围，后生成岩改造作用是形成鲕滩储层的必要条件。更多还原

【Abstract】 Reservoir prediction is a basic work to build an accurate geological model ofreservoir, to accurately estimate reserve, to determine proper development scheme, itis not only applicable in oil and gas exploration, but also in guiding development ofthe oil and gas reservoirs, especially complicated subtle reservoirs or lithologicreservoirs, and has become a key scientific issue concerned by both the academic andindustrial circles domestic and abroad. It is difficult to accurately describe thehorizontal distribution of reservoirs parameters by inter-well correlation orinterpolation, the inversion and seismic attribute techniques which joint logging andseismic information are important reservoir prediction method, have been extended tothe oil and gas development stage, become important component of reservoirdescription technology.The essence of inversion technique is to find the solution of a nonlinear objectivefunction. However, the traditional linear inversion methods sometimes plunge intolocal optimum, or depend on the selection of the initial model, thus affecting thereliability of inversion. Intelligent computation methods such as genetic algorithm orsimulated annealing etc. provide a new way to solve reservoir parameters nonlinearinversion problem, but with shortcomings such as converging slowly or falling intolocal optimum, especially for complicated inversion problems. Neural network hasalso been used to build the nonlinear relationship between seismic attributes andreservoir parameters, and predict reservoir parameters. The essence of error backpropagation learning algorithm of neural network is a gradient method, whichconverges slowly to desirable solution, or may be trapped into local optimum.In this dissertation, I propose a number of new ideas to improve the efficiencyand the optimizing performance of nonlinear inversion, the efficiency of neuralnetwork’s learning algorithm for reservoir parameters prediction. I have studiedreservoir parameters nonlinear inversion method based on improved particle swarmoptimization, and new learning algorithm of neural network which hybrids gradient method and particle swarm optimization, in order to predict reservoir parametersthrough multi-attributes. At the same time, I applied the new methods to predictoolitic reservoir, Feixianguan formation, DCH structure of Eastern Sichuan Basin, andobtain the following results:1, in-depth study of the principles of particle swarm optimization, convergenceconditions, parameters setting etc. basic theoretical problem, particle swarmoptimization algorithm updates the solution more purposely, converges more rapidly.A new chaotic inertia weight adjustment strategy is proposed to further improve thealgorithm’s convergence speed. At the same time, I built the multi-parent crossoveroperator and adopt the group hill climbing idea to improve the performance of particleswarm optimization algorithm. On these bases, I develop new reservoir parametersnonlinear inversion methods based on particle swarm optimization algorithm, and themodel simulations and practical data processing results demonstrate that the newmethod is highly efficient with much improved performance over the traditionalnon-linear methods.2, in-depth study of the error back propagation network, radial basis functionnetwork and their learning algorithm. I have applied the particle swarm optimizationas the train algorithm of back-propagation network and radial basis function networks.Fusion gradient method and the particle swarm optimization are developed to form anew hybrid learning algorithm of radial basis function network, and the accuracy andefficiency of new hybrid algorithm have been greatly improved. On this basis, Ideveloped a new reservoir parameters prediction technique by seismic multi-attributewhich based on hybrid learning algorithm neural network, and applied the techniqueto predict porosity parameters which indicates that the hybrid learning algorithm hasshort training time and high efficiency.3, applying new reservoir parameters nonlinear inversion method based onparticle swarm optimization and new reservoir parameters prediction technique ofhybrid learning algorithm neural network to predict oolitic reservoir, Feixianguanformation, DCH structure of Eastern Sichuan Basin. I have drawn the prediction mapfor reservoir thickness, porosity, and storage coefficient. Based on the comprehensivegeological, seismic and log information conclude that the sedimentary facies controlsoolitic reservoir’s level and scope, and the anadiagenesis is the necessary conditionsfor oolitic reservoir development.更多还原