【作者】 关学忠；

【导师】 段玉波；

【作者基本信息】 大庆石油学院 ， 油气信息与控制， 2009， 博士

【摘要】 人工神经网络(NN)、遗传算法(GA)和模拟退火算法(SAA)的研究是当代信息科学的前沿和热点,也为非线性系统辨识理论的发展开辟了崭新的途径。文中围绕非线性系统辨识和预测方法及其在油田系统的应用,展开了深入研究,提出了基于混合算法的油田系统NN辨识和预测的一系列新方法。主要完成了如下工作。首先,以油田试井解释为应用背景,以试井偏微分方程为依托,经过数学推导,给出了试井解释基函数神经网络(WTBFNN)模型,为试井系统辨识和预测奠定了理论基础。试井解释基函数是典型的复杂多峰函数。它的参数(地层参数)值是试井解释的依据,因而要求其估值应具有唯一性。为了解决这一难题,文中将系统辨识、GA、聚类算法等多项技术融于一体,提出了两种新型混合GA:种族遗传进化算法(SGEA)和启发式GA。仿真实验证明了SGEA的全局收敛性优于带共享机制的GA。文中以马尔可夫链为工具,证明了上述两种新型GA具有全局收敛性。在此基础上,提出了基于新型混合GA的WTBFNN辨识和预测新方法,其中以WTBFNN为模型框架,用F检验法确定模型结构参数,用最小二乘法辨识WTBFNN的权值,用上述两种新型混合GA之一辨识WTBFNN中的地层参数。新方案成功地用于低渗透油田试井问题,取得很好结果。对多组油井实测数据拟合和预测的平均相对误差均在1%以内,同时均得到了对应油井地层参数的全局最优估值,为试井解释提供了科学依据。并且新方案比传统方案显著减少了关井时间,因而具有重大的经济和社会效益。其次,为解决深层火山岩储层预测问题,提出了两种NN辨识器。用径向基函数神经网络(RBFNN)作第1种辨识器模型。用文中提出的新型混合算法对RBFNN的结构、参数和权值进行全面辨识。该混合算法由正交最小二乘法、带惯性项的梯度法和文中提出的优选模糊C均值聚类法组成。用多层前向网络作第2种辨识器模型。用文中所给出的混合SAA训练网络模型。该混合算法由Powell算法和文中给出的自适应SAA组成。由两种辨识器导出的预测器成功地用于火山岩储层预测,具有很高的精度,并且在收敛速度方面优于传统NN预测方案。最后,文中提出一种带Laplace隶属函数的新型模糊NN辨识器和预测器。并且应用微分中值定理和Weiestrass定理证明了该网络具有通用逼近性。在该方案中,采用文中提出的优选K均值聚类法辨识该网络的结构,用所给出的混合算法辨识该网络的前件参数和权值。仿真实验表明新方案优于传统模糊NN预测方案。将上述新方案用于油田积累产量预测,取得很好的结果,预测的平均相对误差在1%以内。特别,实现了对三个采油厂的不同类油井:基础油井,一次和二次加密油井累积产量的长期预测。这些预测结果,为是否继续打加密井的生产决策提供了科学依据。更多还原

【Abstract】 The research of artificial neural networks(NN), genetic algorithms(GA), and simulated annealing algorithms(SAA) are the focus of modern information technology. It also opens a new route for the develop of nonlinear system identification. In this paper, we have made a thorough research in methods of nonlinear system identification and prediction of oil field systems. A set of NN identification and prediction methods can be obtained for oil field systems based on hybrid algorithms. We have finished the research as follows.First, the oil field well test is taken an application example. On the bases of partial differential equations on well test interpretation,the model of well test interpretation basis function neural network(WTBFNN) has been given by mathematical induction, which is the theoretical bases of identification and prediction for well test systems. The well test interpretation basis function is a typical complex multimodal function. Its (stratigrsphic) parameters are the foundation of well test interpretation, so we request their unique estimate values. In order to solve the difficult problem, two new pattern hybrid GA have been put forward by means of technique synthesis for system identification ,GA and cluster algorithms and so on. The two algorithms are species genetic evolution algorithm(SGEA) and heuristic GA. It is shows from simulation experiment that SGEA has advantage over the GA with sharing on global convergence aspect. The global convergence of the two new GA have been proved by using Markovian chain theory. As a result, the new methods of WTBFNN identification and prediction can be got based on the above new hybrid GA. Where WTBFNN is taken as the model frame, the model structure parameters are determined by using F-test method, weight values of WTBFNN are identified by using least-squares(LS) method, and stratigraphic parameters in WTBFNN are estimated by using one of the above two hybrid algorithms. The new scheme can successful be applied to well test interpretation of low penetrability oil field, and excellent results be got. The average relative errors of fitting and predicting are within 1% for many set of oil well data. The global optimum estimate values of stratigraphic parameters can be obtained, which provides scientific basis for well test interpretation. And the new scheme can obviously reduce closing well time than normal well test schemes. Therefore it could bring great economic and social benefit. Second, in order to solve the problem of the prediction of deep volcanic rocks reservoir, the two NN identifiers can be given. radial basis function neural network(RBFNN) is taken as the first identifier model. The structure, parameters and weight values of RBFNN can totally be identified by using a new hybrid algorithm proposed in this paper. The hybrid algorithm is consists of orthogonal LS method, gradient method with inertia terms and optimal fuzzy C means cluster method posed in this paper. The multilayer feedforward NN is taken as second identifier model. The network model can be trained by using a new hybrid SAA given in this paper. The hybrid algorithm is consists of Powell algorithm and adaptive SAA posed in this paper. The high prediction precision is obtained, when the two new schemes are applied to the prediction of volcanic rocks reservoir. And they have advantage over the tradition NN prediction schemes on convergent speed. Last, a novel fuzzy NN identifier with Laplace membership function is posed by us in this paper. The universal approximation of the network can be proved by using differential median theorem and Weierstrass theorem. In the scheme, the structure of the network is determined by using the optimal K means cluster method posed in this paper, the premise parameters and the weight values are identified by the hybrid algorithms given in this paper. The very good results can be gained, when the above new scheme is applied to prediction of accumulative oil product. The prediction average relative errors are within 1%.Specially, the accumulative oil product long-term prediction of basis wells, first and twice thicken wells for three oil recovery plants can be done, and there prediction results can provide the scientific basis for production decision whether or not to drill new thicken wells.更多还原