基于神经网络的系统辨识与控制

发布时间：2018-05-10 21:43

本文选题：非线性系统 + 神经网络辨识　；参考：《华北电力大学(北京)》2017年硕士论文

【摘要】：非线性系统的建模与控制问题一直以来都是控制领域研究的重点与难点。神经网络以其强大的非线性逼近能力,被广泛应用在系统的辨识建模、控制、优化等问题中。同时,很多实际系统中都存在随机噪声且都无法避免它的影响。因此,含有噪声系统的辨识与控制问题成为了神经网络辨识与控制中的一个难点问题。前馈神经网络结构简单、可塑性强,在函数逼近、模式识别、信息分类及数据压缩等领域得到广泛的应用。本文重点介绍了反向传播(BP)神经网络的结构及学习算法及BP神经网络在自适应控制中的应用。但由于前馈神经网络在处理非线性动态系统的辨识与控制中存在一定的缺陷,所以本文也研究了递归神经网络的结构,并重点介绍了离散状态空间神经网络的结构及学习算法,并把它用于非线性系统的辨识中。另一方面,误差纠正学习算法在神经网络的训练学习中广为应用。训练准则一般都是误差的函数,传统的最小均方误差准则仅考虑误差的二阶矩,在非线性或非高斯情形下,其训练性能变得不尽人意。信息测度能够很好的刻画随机变量的不确定性,因此本文详细介绍了熵、信息势、生存信息势等新的测度,并把它们应用于神经网络的辨识与控制中。最后,对于有机朗肯循环(ORC)系统中的过热度控制问题,应用了基于神经网络的自适应控制方法。在实际的ORC过程中烟气的品质将会对蒸发器出口的温度产生影响,而影响烟气品质的烟气的入口温度、烟气流量等扰动因素未必服从高斯分布。本文提出了基于生存信息势的神经网络辨识与神经网络控制,从而保证跟踪误差的随机性最小。
[Abstract]:Modeling and control of nonlinear systems has always been the focus and difficulty in the field of control. Neural network is widely used in system identification modeling, control and optimization for its powerful nonlinear approximation ability. At the same time, many real systems have random noise and can not avoid its influence. Therefore, the identification and control of noisy systems has become a difficult problem in neural network identification and control. The feedforward neural network structure is simple and flexible, and it is widely used in the fields of function approximation, pattern recognition, information classification and data compression. In this paper, the structure and learning algorithm of BP neural network and the application of BP neural network in adaptive control are introduced. However, there are some defects in the identification and control of nonlinear dynamic systems by feedforward neural networks, so the structure of recurrent neural networks is also studied, and the structure and learning algorithm of discrete state space neural networks are introduced. It is applied to the identification of nonlinear systems. On the other hand, error correction learning algorithm is widely used in training and learning of neural networks. The training criterion is generally a function of error. The traditional minimum mean square error criterion only considers the second moment of error. In the case of nonlinear or non-Gao Si, the training performance becomes unsatisfactory. Information measure can well describe the uncertainty of random variables, so this paper introduces some new measures, such as entropy, information potential, survival information potential and so on, and applies them to the identification and control of neural networks. Finally, the adaptive control method based on neural network is applied to the superheat control problem in the organic Rankine circulatory control (ORC) system. In the actual ORC process, the quality of flue gas will affect the outlet temperature of evaporator, while the inlet temperature and flow rate of flue gas, which affect the quality of flue gas, may not be distributed from Gao Si. In this paper, neural network identification and neural network control based on survival information potential are proposed to ensure the minimum randomness of tracking error.
【学位授予单位】：华北电力大学(北京)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP183;TP273

【相似文献】