神经网络技术在预测控制中的应用研究

【作者】 禹柳飞；

【导师】 王辉；

【作者基本信息】 湖南大学 ， 控制理论与控制工程， 2003， 硕士

【摘要】 随着工业控制要求的提高及控制理论与计算机技术的发展，产生了控制效果好、鲁棒性强，适用于控制不易建立精确数学模型且比较复杂的工业过程的预测控制算法，并已在石油、化工、冶金、机械等工业部门的控制系统中得到了成功的应用，是一类很有发展前途的新型计算机控制算法。 本文首先讨论了动态矩阵预测控制基本结构和原理，深入分析了动态矩阵预测控制的预测模型、反馈校正与滚动优化方法、内模控制结构及其稳定性和鲁棒性，并仿真研究证实了动态矩阵预测控制算法是一种先进的控制算法，同时也指出了这种常规预测控制算法面临的困难及存在的问题。在此基础上本文研究了基于神经网络辨识的动态矩阵预测控制新方法，其实质是利用作为对象辨识模型的神经网络产生预测信号，用优化算法求出控制律，从而实现对非线性时变系统的预测控制，神经网络选用具有良好的函数逼近能力的BP网络和RBF网络。首先对被控对象进行离线辨识，在模型辨识达到一定的精度后，再在线递推得到预测模型，最后通过极小化性能指标得到最优控制律。该算法不仅解决了非线性时变对象难以建模的问题，而且还减少了控制器的计算工作量，有利于系统的实时应用。最后介绍了加氢裂化装置的生产工艺流程，建立了基于BF网络与RBF网络的加氢裂化航煤干点的预测模型，并提出了航煤干点基于神经网络辨识的动态矩阵预测控制方案，为实现航煤干点的在线质量控制打下了基础。更多还原

【Abstract】 Along with advancement of industrial control demand, development of control theory and computer technology, a predictive control algorithm is produced with effective control and strong robustness, It is applicable to complex industrial processes and the control system that is not easily to establish the accurate mathematics model, and is successful applied in petroleum, chemical industry, metallurgy and mechanism, and have a good prospect in application.In this paper, it makes a discussion on the basic structure and theory of dynamic matrix predictive control, makes a detailed analysis including its predictive model, its methods of revising feedback and receding horizon optimization, its structure of inside model control (IMC) and its stability & rebustness. Simulation results confirm advanced of the dynamic matrix control algorithm. On the basis of pointing out the problem of the present difficulty and actuality, we propose the idea that DMC combined by neural networks. Actually it uses the neural networks as the identified model of control plant to produce predictive signal, the control law is solved by optimized algorithm. Accordingly we realize the predictive control of the nonlinear and time-variety system. We choose BP and RBF neural networks as identified model for they can approach the function very well. First we identify the controlled plant offline, when the precision reaches a certain extent, we will achieve recursive predictive model by on-line identification. Finally we acquire the optimized control law by minimizing the function of performance guideline. This algorithm not only solves the problem that nonlinear and time-variety plant is difficulty to build model, but also decreases the controller calculative burden. It benefits its application to real-time system. Therefore, its application scope of predictive control is further broadened. In the end, on the basis of introducing technical flow in hydrocracking units, two predictive models of jet fuel endpoint in hydrocracking units are built based on BPNN and RBFN, a new DMC project using the neural networks asthe identified model is proposed, and it provides good conditions foronline quality control of jet fuel endpoint.更多还原