【作者】 沈曙光；

【导师】 王广军；

【作者基本信息】 重庆大学 ， 动力工程及工程热物理， 2009， 博士

【摘要】 系统逆动力学问题是一类典型的动态系统反演课题,其核心内容在于,根据所期望的系统输出过程和已知的系统当前和历史状态,确定应该施加于系统的输入过程。热力系统逆动力学问题已经成为热工过程控制与故障诊断等许多相关领域的一个基础性的重要课题,对该问题进行深入研究,具有十分重要的科学意义和工程实际意义。支持向量机(SVM)是建立在统计学习理论基础上的一种机器学习算法,具有理论完备、全局优化、适用于小样本建模和泛化能力强等优点,已经成为复杂系统辨识的有效工具。本文研究了基于SVM的热工过程逆动力学模型辨识与控制问题,主要研究工作与结果包括以下五个方面。①针对两类典型热工对象,建立了逆动力学模型的基本结构;提出了一种基于对象动态过程相关分析的多输入多输出(MIMO)系统逆动力学模型输入向量的约简方法,通过对象动态过程的相关分析,获得对象逆动力学过程的最大相关量和主导参数,在此基础上确定MIMO系统逆动力学模型的输入向量。利用该方法对双输入双输出系统逆动力学模型进行了深入研究,提出了双输入双输出系统逆动力学模型的四种基本结构。②与标准SVM相比,最小二乘支持向量机(LSSVM)将SVM的学习问题转化成线性方程组的求解问题,在函数估计和逼近中得到了广泛应用,其主要问题在于,在基于LSSVM的系统建模过程中,模型的每次更新都需要进行一次矩阵求逆,难以用于复杂系统的在线辨识。本文将递推最小二乘法(RLS)和LSSVM算法相结合,利用RLS在线调整LSSVM的权向量和偏移量,通过修剪算法减少支持向量个数,提出了一种递推最小二乘支持向量机(RLSSVM),明显地提高了系统模型在线辨识过程的实时性和模型的精确性。文中利用RLSSVM实现了几类典型热工对象逆动力学模型的在线辨识,验证了基于RLSSVM的系统逆动力学过程模型在线辨识方法的有效性。③针对LSSVM方法存在的核函数及算法参数选取问题,以及系统模糊辨识模型泛化能力不足问题,将LSSVM方法和模糊辨识方法相结合,提出了一种基于最小二乘支持向量机的模糊辨识方法(FLSSVM),并证明了该方法与LSSVM的等价性;在基于FLSSVM模型在线辨识算法中,采用滚动时间窗概念,由模糊竞争学习算法修正聚类中心;在结论参数辨识中,对滚动时间窗内的样本数据根据采样时间的不同,分配不同的权值。基于FLSSVM的热工对象逆动力学过程辨识结果表明,该方法具有较高的辨识精度和良好的跟踪性能,具有良好的泛化能力和对噪声的适应能力。④将前述的逆动力学模型辨识方法与自适应逆控制技术相结合,建立了一种基于FLSSVM逆动力学模型的自适应逆控制系统。在控制过程中,通过FLSSVM方法在线辨识控制对象的逆动力学模型,对控制器进行在线更新;同时利用基于FLSSVM的逆动力学模型自适应地完成系统输出扰动消除过程。针对火电机组过热汽温对象和负荷对象设计了相应的自适应逆控制系统,控制过程仿真实验表明,所建立的自适应逆控制系统具有良好的控制性能和自适应能力,并表现了良好的输出扰动消除效果。⑤根据直流锅炉汽温对象典型动态特性的分析结果,明确了现有的汽温分段控制方法存在的主要问题,建立了一种基于逆动力学模型的直流锅炉给水及过热汽温的自适应逆控制系统,通过逆动力学模型在线辨识,同时考虑直流锅炉出口汽温以及微过热汽温等对给水量和各级喷水量的共同需要,实现直流锅炉给水及过热汽温的综合控制。仿真实验表明,该控制系统具有良好的控制品质和自适应能力,并有效地消除了现有的汽温分段控制方法中存在的控制量反复振荡现象。更多还原

【Abstract】 System inverse dynamics is a class of typical inverse subject, its key problem is that the input of system is determined according to the anticipant output and known current and past information. Inverse dynamics of thermal system has become an important issue in many relative areas such as thermal process control and fault diagnosis. It has very important scientific and practical significance that study on inverse dynamics of thermal system.Support vector machines(SVM) based on the statistical learning theory is a new approach in machine learning. SVM has become an effective tool for complexity system identification because of its advantages such as firm mathematic theory foundation, strict theory analysis, global optimization as well as good adaptability and generalization. Inverse dynamic model identification and control for thermal system based on SVM is studied in this dissertation. The main works include the following five parts:①The basic structure of inverse dynamic model is built for two typical thermal system. A reduction method for input vector of multiple input and multiple output(MIMO) inverse dynamic model is proposed. The largest relation value and leading parameters are obtained through analyzing dynamic characteristics of object, and then the input vector of MIMO system inverse dynamic model is determined. Four basic structures of input vectors are proposed by studying on inverse dynamic model of dual input and dual output system.②Compared with SVM, LSSVM whose learning algorithm is changed to solve linear equations has widely applied in function estimation and approximation. However, LSSVM is difficult to realize on-line identification for complexity system because it need solve inverse matrix when updating model in dynamic modeling process. Recursive least square support vector machines(RLSSVM) is developed by combining recursive least square(RLS) algorithm with LSSVM. The parameters of model are updated by RLS algorithm. To speed up the computational speed, the number of support vectors is reduced by doing pruning. The accuracy of model and real-time of on-line identification is improved in this method. On-line identification for thermal system inverse dynamic model is realized by RLSSVM. The simulation results show the effectiveness of on-line identification method which is based on RLSSVM for inverse dynamic model.③Because selecting kernel function and its parameters is very complex in LSSVM and T-S fuzzy model has weak generalization performance. Fuzzy modeling method based on LSSVM(FLSSVM) is proposed by combing LSSVM with fuzzy model. FLSSVM is proved that is equivalence with LSSVM. The concept of sliding time widow is adopted in on-line algorithm of FLSSVM, the cluster centers are updated by fuzzy competitive learning algorithm. Different weighs are assigned to the samples in sliding time window according to the sampling time when identifying conclusion parameters. The simulation results of identification for thermal inverse dynamic model show that modeling method based on FLSSVM has good precision and tractability. The model based on FLSSVM also has good generalization ability and adaptability to noise.④An adaptive inverse control system based on inversed dynamic model of FLSSVM is constructed by combining the identification method of inverse dynamic model with adaptive inverse control method. The inverse dynamic model of control object is identified on-line by using FLSSVM, and then the controller is updated in the control process. The output disturbance of system is canceling adaptively with inverse dynamic model based on FLSSVM. Adaptive inverse control systems are designed for superheated steam temperature and unit load of power plant. Simulation results show that the adaptive inverse control system has good control performance and adaptability, and has good disturbance canceling effect.⑤The main problem of present two section control method is ensured according to analyze typical dynamic characteristics of steam temperature in once-through boiler. An adaptive inverse control system is built for feed water and superheated steam temperature in once-through boiler based on inverse dynamic model. The control of feed water and superheated steam temperature is realized through on-line identification for inverse dynamic model. In this control system, the demand of feed water and spray water flow to superheated steam temperature and micro-superheated steam temperature is considered. Simulation results show that the control system is designed has good control performance and adaptability and can void the repeated oscillation phenomena of control variables which appear in present two section control system.更多还原