基于二次逼近神经网络的反应釜预测控制

Predictive Control of Reactor Based on Quadratic Approximation Neural Network

摘要:

针对在化工生产过程中使用连续搅拌反应釜(Continuous Stirred Tank Reactor,CSTR)时存在的控制方式不便,调节的精确度不高等问题;在对实际问题进行分析建模的基础上,提出了一种基于二次逼近神经网络模型的预测控制方法;该方法首先利用多层前馈神经网络模型去逼近连续搅拌反应釜系统的多步预测值,其次在已创建的预测模型的基础上优化并求解预测控制的二次目标函数,以得到最优的控制参数,最后由通过泰勒展开式的二次逼近得到非线性预测控制器的最优解;通过对控制模型的模拟以及带入相关参数进行仿真实验,对连续搅拌反应釜控制系统的仿真结果进行分析表明:该方法控制精确度较高,并且是可行有效的,能够使生产效率得到显著提高且保证了产品的质量,具有较高的实用价值。

Abstract:

A predictive control method based on neural network is presented for the problem of production inconvenience and low precision in a continuous stirred tank reactor(CSTR) which is widely used in chemical process. Firstly, using a multi-layer feed forward neural network to approximate the CSTR system’s multi-step predictive value, secondly based on the forecast model for solving quadratic objective function optimization predictive control by the Taylor expansion approximation quadratic nonlinear predictive control law controller. The simulation results of CSTR show that :that essay’s method has high accuracy and that method is feasible and effective. The utility model has the advantages of high production efficiency, high product quality and high practical value.