【作者】 李奇安；

【导师】 王树青；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2005， 博士

【摘要】 生产实践发展的需要催生了模型预测控制方法,模型预测控制理论发展的目的是更好地服务于生产实践。因此,如何把先进的模型预测控制算法应用于实际控制中是模型预测控制发展的根本性问题之一。论文正是从这一观念出发,在前人研究的基础上,对广义预测控制算法提出了一系列的简化实现方法。论文包括以下内容: 首先,介绍了模型预测控制方法的发展过程、现状、目前存在的局限性以及发展趋势。分析了广义预测控制算法的特点及应用时存在的困难。在总结模型预测控制方法常用简化实现方法和广义预测控制算法简化实现已取得的成果的基础上,提出论文所要进行的研究工作。 第二,对于物理可实现的多变量系统,其CARIMA模型的A(z^-1)与C(z^-1)多项式矩阵总可以构造成对角形式,从而可以给出这种模型的广义预测控制算法的完整求解过程。在算法的推导过程中,显式地考虑了过程纯滞后项,以提高计算效率。利用这种形式的模型结构,不但广义预测控制算法的求解过程可以得到很大程度的简化,而且相应的模型参数辨识问题也得到了简化,可以把一个多输入多输出模型的大型参数辨识问题分解成多个多输入单输出模型的小型参数辨识问题。 第三,通过对模型预测输出自由响应项的进一步分析,得到了状态反馈结构形式的广义预测控制器,控制增量等于控制器系数与设定值、过程输入输出历史数据的乘积,控制器系数只与模型参数和设计参数有关,控制器系数维数由预测时域与模型结构参数决定。消除了在非自适应模式下在线求解模型输出自由响应的必要,可以像PID控制器一样实现广义预测控制器。 第四,利用CARIMA模型直接递推得到了更加简洁的广义预测控制器,控制增量等于控制器系数与设定值、过程输入输出历史数据、模型预测误差历史数据的乘积,控制器系数只与模型参数和设计参数有关,控制器系数维数只由模型结构参数决定。在自适应模式下无需进行Diophantine方程的求解,在非自适应模式下将实现难度与计算量降到了最低。 第五,通过对多变量广义预测控制算法内在机理的分析,指出多变量广义预测控制算法实际上是一种函数映射,是从模型参数空间到控制器系数空间的映射,并利用神经网络的映射能力来实现多变量广义预测控制器系数的快速计算,更多还原

【Abstract】 It is the development of industrial process control that has speeded up the emerging of model predictive control (MPC) methodology and the basic motivation of using MPC technology is to ensure significant economic benefits. Thus, one of the most essential problems of MPC technology development is how to implement these advanced algorithms in real world effectively. Just from this view, a set of simplified implementation for generalized predictive control (GPC) are presented based on the current framework of MPC. The dissertation is organized as follows:First, a brief history of MPC technology development is presented, followed by its current features and limitations of existing technology and its future trends. The distinguishing features of GPC approach and its application difficulty are discussed. After summarizing the general ways to simplify the implementation of MPC and the achievements in simplifying implementation for GPC, the main research works are given.Second, for most of the physical realizable processes, the matrices C(z-1) andA(z-1) of their Controlled Autoregressive Integrated Moving Average (CARIMA)model can always be diagonally constructed, so that the formulation of GPC can be developed in more detail while explicitly considering the dead time in order to improve the computational efficiency. This model structure greatly simplifies not only the development of the GPC but also its parameter identification which can be transformed into a set of multiple input single output model parameter identification problems.Third, a state-feedback like controller of GPC is obtained by further manipulating the free response of the output predictor, whose control increment equals to the controller’s coefficients multiplied by set-points and historical plant input and output data. The controller’s coefficients are only determined by the model parameters and design parameters and its dimension is determined by the model structure parameters and predictive horizon, which eliminates the need to compute the free response on-line and makes the implementation of GPC controller as easy as thatof PID under the non-adaptive mode.Fourth, a more concise GPC controller is obtained by directly manipulating the output predictor using the multivariable CARIMA model recursively, whose control moves are the product of the controller’s coefficients and set-points, historical input/output data of the plant and predictive errors of the predictor. The controller’s coefficients are determined only by the model parameters and design parameters and its dimension only depends on the orders of the model, which avoids solving Diophantine equations on-line under adaptive mode and reduces difficulties of implementing the GPC controller and the computational overhead to the lowest limit under the non-adaptive mode.Fifth, it is pointed out that the multivariable GPC algorithm is essentially a kind of functional mapping from the multivariable process model parameters’ space to the multivariable GPC controller’s coefficients’ space by analyzing its intrinsic mechanism. This mapping can be realized by BP neural network to obtain the GPC controller’s coefficients from the model parameters directly, which can extremely reduce the computational overhead on-line and simplify the implementation of the GPC controller.Sixth, the above schemes developed in this dissertation are compared by a set of contrast experiments on a nonlinear liquid level equipment. Their feasibility, validity and equivalency are demonstrated by experiment results.Last, a summary is given to show what has been done in this paper. The applicable scope of these methods developed in this dissertation is discussed, followed by some items that must pay attention to when implement these methods and future potential research opportunities.更多还原