预测控制是七十年代末产生并发展起来的一类新型计算机控制算法。因其控制性能良好、易于实现、鲁棒性好等优点，在工业过程控制中得到广泛应用。目前，虽然线性单变量系统的预测控制理论发展较为成熟，但对强非线性系统，线性预测控制方法已不能满足要求，在此情况下就需要采用非线性预测控制方法。由于非线性预测控制自身的复杂性，无论是理论分析还是实际应用都相当困难，因此研究非线性预测控制方法，具有重要意义。本文针对工业过程的实际问题对非线性预测控制进行了较深入的研究。 本文首先在广义预测控制（GPC）基础上，针对模型中随机干扰项系数无法在线估计的问题，研究了一种引入滤波器的GPC，给出了具体的算法并通过仿真总结出滤波器的作用，并对GPC的参数选取进行了定性的理论分析，讨论了主要参数的选取方法，使其更加完善。 其次，从非线性系统的结构和特点出发，研究了两种适合于非线性预测控制的滚动优化方法，一种是两步法，该方法采用Hammerstein模型，将非线性预测控制问题分解为线性模型的动态优化和非线性模型的静态求根问题，第一步对线性子系统应用GPC得到一个所期望的中间变量，第二步由中间变量通过求解非线性方程得到实际的控制量，对非线性方程无实根的情况，进行了两种近似方法的探讨。另一种是整体求解方法，该方法采用遗传算法实现滚动优化，直接对控制量编码，将非线性部分纳入性能指标并以此为目标函数，直接求控制律，算法中通过提高初始种群质量来解决实时性问题。 最后，通过几个典型实例的仿真，验证了上述算法的可行性及提高控制性能措施的有效性，为其应用于工业现场控制奠定了基础。

Predictive control is a class of digital control algorithms developed from industrial process in the late 1970's. The popularity of these methods is due to the facts that they can offer good performance, be understood and formulated easily and robustness to the model uncertainty. So far, the theory of linear predictive control has matured considerably, but it dose not apply to the strong nonlinear objects where nonlinear MPC is required. Nonlinear MPC is rather difficult both in theory and in application because of its complexity. Therefore, the study of nonlinear MPC algorithm is of great significance. Based on the factual problems in industrial process, nonlinear predictive control is researched profoundly in this paper. Firstly, a primary algorithm theory of generalized predictive control is introduced. In order to solve the problem that the coefficient of random disturbing item can't be estimated on-line. A kind of Generalized Predictive Control (GPC) is researched, and filter is adopted in GPC. The specific algorithm is put forward, and the function of the filter is summarized by simulating GPC. Then the selection of GPC parameters is analyzed rationally in the paper, and the method of selecting main parameters is discussed to make GPC more perfect. Secondly, based on non-linear system’s structure and characteristics, two methods of moving horizon optimizations are put forward, both of the methods are suitable for non-linear predictive control. One is the method of two-step, and Hammerstein model is adopted in the method. In this strategy, MPC problem is divided into a dynamic optimization problem upon linear model and a static rooting problem of nonlinear algebraic equation. The Generalized Predictive Control algorithm is firstly applied to the linear part of the model, and then the required control variable is obtained by extracting the root of the nonlinear equation. Based on the non-linear equation without real root, two approximate calculation methods are offered, and two methods are compared with each other. The other is the method of entire solution; genetic algorithm is adopted in the method. In the algorithm, controlling variables are coded directly nonlinear part is included in performance indexes as a target function to solve the control law. Real-time problem is solved by improving the quality of initial stock in algorithm. Through the simulation of several typical examples, the feasibility of the nonlinear MPC algorithms and the correlative improved measures presented in this paper are proved, and the advantages and disadvantages of these algorithms are compared according to the results of simulation. And it lays the foundation of using in the industrial field control.