【作者】 刘霞；

【导师】 李大字；

【作者基本信息】 北京化工大学 ， 控制理论与控制工程， 2008， 硕士

【摘要】 智能预测控制是针对复杂的受控系统,采用某种智能模型与典型的预测控制算法相结合构成的一类智能型预测控制系统,它弥补了传统预测控制算法精度不高、仅适用于线性系统、缺乏自学习和自组织功能、鲁棒性不强的缺陷。智能预测控制可以处理非线性、多目标、约束条件等异常情况。因此,智能预测控制是当前预测控制研究的热点之一。本文主要研究一种新型神经网络即一般化学习网络(Universal Learning Network)在非线性系统辨识以及预测控制中的应用。一般化学习网络具有节点之间有多重分支、任意两个节点互连且节点之间可具有任意的时间延迟的特点,因此能够应用在高度非线性复杂系统的辨识中。本文首先详细介绍了一般化学习网络的结构以及学习算法后,将该网络应用于对机器人手臂的实测信号进行系统辨识,通过系统仿真验证了具有多重分支的一般化学习网络优于单分支的神经网络。然后分别用一般化学习网络和常规的递归神经网络Elman对多变量连续釜式反应器(CSTR)进行系统辨识比较,仿真结果进一步验证了一般化学习网络结构比递归神经网络Elman的辨识精度更高,且网络结构更简洁紧凑。基于一般化学习网络对复杂系统良好的辨识能力,本文提出了将一般化学习网络应用于非线性系统预测控制中,利用一般化学习网络构建一个一般化学习网络预测模型(ULNP)来预测模型未来时刻的输出值,然后利用BP神经网络控制器(NNC)实现基于模型的预测控制。本文详细描述了一般化学习网络预测模型和神经网络预测控制器模型及其在线学习的推导过程。仿真结果证实了所提出算法能够提供满意的跟踪性能。更多还原

【Abstract】 Intelligent predictive control aim at complex system control using an intelligent model combined with typical predictive control algorithm to constitute a class of intelligent predictive control system, which makes up the defects of traditional control algorithms, such as low accuracy, only adaptive to linear System, lack of self-learning and self-organization functions, and tso on. Intelligent predictive control can handle non-linear, multi-objective, constraints and other anomalies. Therefore, intelligent predictive control is one of the hot spots in current researches for predictive control. The research relates to the identification and the predictive control of nonlinear system based on a new type of neural network that is Universal Learning Network (ULN).Universal Learning Networks consist of a number of nodes and branches for inter-connecting the nodes and each pair of nodes can be connected to each other by multiple branches with arbitrary time delays. With all these structural characteristics, it can be used in modeling the highly complicated nonlinear system. Firstly a detail introduction of the structure and learning algorithm of Universal Learning Network was given in the paper, and then the network was used to model a robot arm’s measured signal system, through simulation we can see that universal learning network with multiple branches is better than the single-branch neural network. In the following part of the paper, both the universal learning network and the conventional recurrent network which is Elman have been used to identify the CSTR system, the simulation results further validate that the universal learning network has higher accuracy than the Elman network when they are used in identification, and furthermore, the network structure is more simple and compact.Based on the good identification ability of Universal Learning Network in complex system, Universal Learning Network was proposed to be used in the predictive control of the nonlinear system, Universal Learning Network was used to build a Universal Learning Network predictor (ULNP) to predict the output value on the next moment, and then BP neural network controller (NNC) was used to realize the model-based predictive control. The modeling and the on-line recursive learning algorithm of the predictor and the controller are explicated in detail. Simulation results show that the proposed control algorithm can give good tracking performance for an illustrative nonlinear system.更多还原