【作者】 任振华；

【导师】 雎刚；

【作者基本信息】 东南大学 ， 能源信息自动化， 2021， 硕士

【摘要】 在热工控制领域,被控对象往往表现出大惯性、非线性等特点,这增加了热工对象模型辨识与优化控制的难度。目前热工过程的模型辨识主要基于动态特性试验,但实验往往具有局限性以及操作难度大等不足,而基于现场数据的热工过程神经网络辨识可以很好规避这些问题。目前火电机组热工过程广泛采用的PID控制难以满足大惯性、非线性对象的控制要求,结合一些先进控制策略对大惯性、非线性过程进行控制就显得很有必要。针对以上问题,论文研究了基于现场数据与神经网络的热工过程模型辨识并与Smith预估控制相结合,提出了一套解决大惯性、非线性热工过程控制问题的方案。主要研究内容及取得的研究成果如下:对BP神经网路的辨识性能从两方面出发进行优化。一是使用变化率误差平方和加传统误差平方和作为神经网络的性能指标,解决传统性能指标进行动态过程神经网络模型辨识时存在的辨识精度和泛化能力相矛盾的问题;二是在新型性能指标基础上加入了神经网络输入节点与隐含层节点的权值系数平方和项,并利用蚁群算法对神经网络进行剪枝操作,进一步提高神经网络的泛化能力。仿真研究表明,使用新型性能指标并剪枝后的神经网络模型,模型结构更为精简、泛化能力更好、准确体现了过程阶次。提出了一种改进的自平衡高阶对象的一阶模型建立方法,该方法参考自平衡高阶对象的阶跃响应曲线建模方法,在该方法基础上提出一个新的一阶模型惯性时间值。该值的引进避免了传统惯性时间值偏大的问题;该值使用较为方便,无需在传统惯性时间值上进行比例调整就可获得合适的惯性时间值。仿真研究结果表明,基于改进惯性时间值的一阶模型比基于传统惯性时间值及其比例值的准确性更高。提出了传统误差平方和加控制器输出斜率平方和这一新型性能评价指标,该指标可以有效避免Smith预估控制在遗传算法PID参数寻优后控制器输出不断波动的问题。仿真研究表明,基于该改进性能指标的PID参数寻优,不仅在Smith预估控制中表现优异,而且在单回路以及串级控制中都能得到不错的寻优结果。提出了一种基于自适应预估器的大惯性、非线性热工过程Smith控制方案。该方案既可以利用Smith预估控制在大惯性过程控制中的良好性能,又可以避免非线性过程特性变化带来的控制问题。与串级控制系统相比,该方案使用的是单回路控制,控制系统的结构更为简化,有利于工程应用。仿真研究表明,该方案有效减小了控制过程的惯性、大大减小了非线性对控制性能的影响,拥有比串级控制更高的控制品质。提出了大惯性、非线性热工过程辨识与控制的应用方案。该方案以某超临界机组过热汽温过程为例,基于Smith预估控制,根据网络模型获取对应负荷、煤质工况点的过热汽温过程特征参数用于实时调整Smith预估器参数,实现了Smith预估器的工况自适应,从而解决了工况变化带来的非线性控制问题,并通过仿真验证了方案有效性。更多还原

【Abstract】 In the field of thermal control,the controlled objects often show the characteristics of large inertia and non-linearity,which increases the difficulty of model identification and optimal control of thermal objects.At present,the model identification of thermal process is mainly based on dynamic characteristic test,but the test is often limited and difficult to operate.The neural network identification of thermal process based on field data can well avoid these problems.At present,the PID control widely used in thermal process of thermal power unit cannot fulfill the requirement of large inertia and non-linear object.It is necessary to control large inertia and non-linear process with some advanced control strategies.To solve the above problems,this paper studies the identification of thermal process model based on field data and neural network,and combines it with Smith predictive control,and presents a set of solutions to large inertia,non-linear thermal process control problems.The main research contents and achievements are as follows:The identification performance of BP neural network is optimized from two aspects.Firstly,using the sum of change rate error squares and traditional error squares as the performance indicators of the neural network,the contradiction between recognition accuracy and generalization ability in dynamic process neural network model identification with traditional performance indicators is resolved.Secondly,on the basis of the new performance indicators,the sum of the square of the weights of the input nodes and the hidden layer nodes of the neural network is added,and the ant colony algorithm is used to prune the neural network to further improve the generalization ability of the neural network.The simulation results show that the structure of the pruned neural network model is simpler,the generalization ability is better,and the process order is accurately reflected by using the new performance index.An improved first-order model building method for self-balanced high-order objects is presented,which refers to the step response curve modeling method for self-balanced high-order objects,and a new first-order model inertia time value based on this method is presented.The introduction of this value avoids the problem that the traditional inertial time value is too large.This value is more convenient to use,and an appropriate value of inertial time can be obtained without scaling the traditional value of inertial time.The simulation results show that the first-order model based on the improved inertial time value is more accurate than that based on the traditional inertial time value and its scale value.A new performance evaluation index,square of traditional error and square of output slope efficiency of PID controller,is presented,which can effectively avoid the problem that Smith predictive control fluctuates continuously after the PID parameters are optimized by genetic algorithm.The simulation results show that the PID parameter based on the improved performance index not only performs well in Smith predictive control,but also obtains good results in single-loop and cascade control.A Smith control scheme for large inertia and non-linear thermal processes based on an adaptive predictor is presented.This scheme can take advantage of the good performance of Smith predictive control in large inertial process control and avoid the control problems caused by the change of non-linear process characteristics.Compared with cascade control system,this scheme uses single loop control,which simplifies the structure of the control system and is conducive to application.The simulation results show that the scheme effectively reduces the inertia of the control process,greatly reduces the influence of non-linearity on the control performance,and has a higher control quality than cascade control.An application scheme for identification and control of large inertia and non-linear thermal processes is presented.This scheme,based on Smith predictive control,takes the superheated steam temperature process of a supercritical unit as an example,obtains the characteristic parameters of the process of superheated steam temperature corresponding to the load and coal quality point according to the network model,which can be used to adjust the parameters of Smith predictor in real time.The adaptive operation of Smith predictor can solve the non-linear control problem caused by the change of working condition.It is demonstrated by simulation that this scheme is effective.更多还原