【作者】 李天舒；

【导师】 李文秀；

【作者基本信息】 哈尔滨工程大学 ， 控制理论与控制工程， 2006， 博士

【摘要】 对于油田生产来说，要保证一个好的经济效益，就必须有一个高的、稳定的产油量。这是油田生产开发的中心任务。油田产量预测是科学管理油田和制定经济计划的依据。目前，各油田已拥有一套完善的包括单井数据在内的油田产量数据管理系统，急需将这些数据利用起来具体地指导油田的实际生产开发。 油田开发是一个复杂的非线性动力学系统，油田产量变化受多种因素控制，导致其表现形式既有确定性又有随机性。总的来说，油田产量预测是一个多因素非线性预测问题。目前所采用的油田产量预测方法的预报相对误差都在10％左右，并且这些方法对油田开发过程的时变性和各种随机干扰因素具有不适应性。因此，有必要借助于其它的分析工具对油田产量进行预测。我们可以得到油田产量的历史数据，可将这些数据看成是时间序列并利用时间序列分析的方法对其进行建模及预测。 本论文以油井产量时间序列为对象，对RLS算法、混沌时间序列性质鉴别方法、混沌时间序列预测方法等进行了深入细致的研究。 论文首先分析RLS算法的性能，在此基础上为了提高收敛速度，提出一种改进的RLS算法，利用改进的RLS算法对某油田油井产量进行了建模及预测。通过对预测结果的分析发现RLS预测方法不能对油井产量进行精确地多步预测。 其次，为了更好的了解油井产量时间序列地性质，利用相空间重构方法重构该序列的吸引子，并计算其维数和最大Lyapunov指数，指出油井产量时间序列具有混沌特征。同时，针对伪近邻算法效率低的缺点，提出一种快速伪近邻法选择嵌入维数的新方法，该方法可将原算法时间复杂度由从O(M(M-1))降低到O(3M)。 再次，对训练支持向量机的序列最小优化方法进行了深入研究，针对原算法选取工作集过于随机的缺点，提出一种基于遗传算法选择工作集的新方法，该方法能够保证每次所选取的工作集使得目标函数变化最大，从而使目标函数尽快向极值收敛。仿真结果表明，该方法可大大加快训练速度；同时，针对以往支持向量机参数选取过于主观的缺点，提出了基于遗传算法方法选择支持向量机参数的方法，仿真结果表明，这种新的参数选择方法可在不明显增加支持向量个数的基础上减小泛化误差。更多还原

【Abstract】 The most import task for oilfield is to ensure a high and stable output. The accuracy of prediction is the basis of oilfield management. Now, most oilfields have held a set of output data including single well data. And they urgently want to use these data for the development of oilfield.In fact, the process of oilfield development is a complicated nonlinear dynamics. In one word, the prediction for oilfield output is nonlinear. At present, the relative error of the existing prediction methods for oilfield output is about 10%, and these methods can not fit to stochastic noises. We have the oilfield output data, so we can use time series analysis tool for modeling and prediction.The research object of this thesis is the time series of some oil wells outputs. The research fields of the thesis include RLS algorithm, the methods for detection chaos in time series, and the methods for time series prediction.First, the characteristics of RLS algorithm are analyzed deeply, and an improved RLS algorithm is proposed. With the proposed algorithm, oil well outputs are predicted. Simulation results show that, RLS algorithm can not make long time prediction.Second, in order to understand the characteristics of time series of oilfield outputs more clearly, phase space reconstruction is used to reconstruct the attractor of the time series. The maximum Lyapunov exponent and the dimension of the attractor show that the time series of oil well outputs are chaotic.Studies show that the method of false neighbors is not effective, so a quick falseneighbor method that can improve the time complexity from O(M(M-1)) to O(3M).To avoid selecting work set of sequential minimal optimization algorithmrandomly, a new method based on GA is proposed. The work set selected by theproposed method can make the change of target function as large as possible.Simulation results show that the proposed method can speed the training process.And to select suitable parameters with support vector machine, a GA-basedmethod for parameter selection is proposed, too. Simulation results show that thisnew method can reduce generalization error with little increase in support vectors.Finally, based on the identification of characteristics of oil well outputs time更多还原