【作者】 陈鹏；

【导师】 吕勇哉；

【作者基本信息】 上海交通大学 ， 控制理论与控制工程， 2011， 博士

【摘要】 以预测控制为典型代表的先进控制技术在过去的几十年中取得了许多重要成果和广泛应用。随着现代工业的发展,对大型、复杂和不确定性系统的控制品质要求不断提高,逐渐涉及到很多具有强非线性、难以用数学模型精确描述的复杂系统建模与实时优化问题,基于线性模型的预测控制的局限性日益明显。非线性预测控制的提出为解决上述问题提供了可行的途径,然而非线性模型的引入也带来了一系列理论与实际难题,如预测模型结构/参数选择、控制变量的滚动优化等。这就促使工业过程领域的学者们致力于适合非线性过程建模、控制及优化的高效率全局优化算法的研究与开发。近年来,进化算法(Evolutionary Algorithm,EA)凭借其求解复杂非线性优化问题所表现出的全局搜索能力、通用性和易用性,在系统分析、模型辨识、控制器设计等领域得到了广泛应用。然而,进化算法存在寻优结果一致性差、局部搜索效率低、求解精度以及实时性不能令人满意等不足。针对上述问题,Memetic算法(Memeticalgorithm, MA)将基于全局的随机搜索和基于局部的启发式搜索有机地结合起来。该类算法采用进化算法的操作流程,引入局部启发式搜索来模拟由大量专业知识支撑的变异过程,在保证较高收敛性能的同时又能获得高质量解,大大提高了算法的搜索效率和求解精度。本研究立足于非线性模型预测控制领域,对预测模型和滚动优化两个关键环节中存在的研究难点及其应用现状做了深入而全面的总结。基于最新提出的统计物理学概念——极值动力学,借鉴协同进化思想,在MEMETIC框架下提出了两种基于极值动力学的高效率混合优化算法,通过一系列典型优化问题验证了所提算法的有效性,并进一步将其应用于非线性预测控制的模型辨识和滚动优化中。本文的主要工作包括:(1)针对现有的极值优化算法在理论基础、局部搜索能力方面存在的不足,对协同进化、Memetic算法以及极值动力学的内在联系进行了分析。基于Memetic算法的框架,将确定性局部搜索算法——列文伯格-马夸尔特算法(Levenberg–Marquardt,LM)作为局部搜索算子引入到极值优化算法(Extremal Optimization, EO)的搜索过程中,提出了一种用于求解无约束优化问题的混合EO-LM算法,并将其应用到典型的无约束优化问题——神经网络的学习中,取得了比传统神经网络学习算法更好的结果。(2)在上述算法的基础上,考虑到预测控制中的优化问题常常带有各种约束,设计了可求解无约束/有约束非线性优化问题的EO-SQP算法,并通过一系列典型非线性数值优化问题的仿真验证了算法的有效性,进一步,结合仿真实例的优化过程对所提算法动力学特点进行了分析。与传统优化算法相比,所提出的EO-SQP算法具有良好的优化性能和求解效率。(3)根据非线性预测控制中对预测模型的需求,提出了一种可对支持向量机模型核函数/参数同时优化的EO-SVR算法,并应用于非线性系统的动态建模;同时,采用可变窗口变异策略对EO-SQP算法进行改进,使之可用于NMPC的滚动优化。在上述预测模型和在线优化算法的基础上,设计NMPC控制器,并通过一类典型的非线性系统——连续混合反应釜的仿真,验证了所提算法的有效性。(4)针对转炉炼钢这一典型的非线性工业生产过程,采用EO-LM算法对转炉炼钢终点质量神经网络预测模型的连接权值和偏置系数进行优化;并提出了一种基于EO-SQP的转炉炼钢配料调整算法。在此基础上,设计开发了转炉炼钢终点质量预测与控制系统,通过实际工业数据验证了本文所提算法的有效性和可行性。更多还原

【Abstract】 As oneofthe mostpopular advancedprocesscontrol (APC) solutions,the model predictive control (MPC) has been widely applied in processindustries during recent years. Due to the rapid development of industryand technology, the control performance requirements for the large-scale,complicated and uncertain system keep rising. Basically, the issuesmentioned above mainly involve solving various types of complicatednonlinear optimization problems. The linear model predictive control(LMPC), which usually relies on a linear dynamic model, is inadequateforabovementionedproblemsandthelimitationbecomesmoreandmoreobvious. Nonlinear model predictive control (NMPC) appears to be aperspective method. However, the introduction of nonlinear predictivemodel brings some difficulties, such as the parameters/structure selectionof the nonlinear predictive model and the online receding horizonoptimization. All these issues encourage researchers and practitioners todevote themselves to the development of novel optimization methodswhich are suitable for the applications in nonlinear model predictivecontrol.The evolutionary algorithms (EAs) have been widely used to solvemanycomplex nonlinear problems in control engineering,suchas systemanalysis, model identification and controller design, and shown theadvantages of global search efficiency, generality and practicality.However, due to the inherited shortcomings of Darwin’s theory, theevolution of computations often suffer from low local search capabilityand accuracy. In order to overcome these limitations,“MemeticAlgorithm”(MA) was proposed by incorporating the modern evolutiontheory and Dawkins’s“Meme”concept into evoluation computation. MAis a very efficient stochastic heuristics for global optimization, bycombining the global search nature of EA with local search to improve individualsolutions.Thisthesisstartswiththegeneralreviewoftwokeyresearchissuesinnonlinear model predictive control: prediction model andrecedinghorizonoptimization. Base on the newly proposed concept called“Extremal dynamics”in statistical physics and co-evolution theory, weproposed two highly efficient optimization algorithms under theframework of MA, and then applied them to a number of benchmarkproblems, such as numerical optimizations, NMPC online optimizaitionand industrial applications. The main topics studied in this thesis aresummarizedasfollows:(1)Regarding the disadvantages of current evolutionary computationin theoretical foundation and local search effectiveness, theinter-relations between co-evolution, memetic algorithm andextremal dynamics are studied. Anovel hybrid memetic algorithmwith the integration of“Extremal Optimization”and“Levenberg–Marquardt”is proposed, for unconstrained nonlinearoptimization. The proposed method is then employed for trainingof multilayer perceptron (MLP) networks and its effectiveness isdemonstrated.(2)Considering the existence of constraints in optimal controlproblems, a novel EO-SQP algorithm is designed for theoptimization of nonlinear problems with/without constraints. Thesearch dynamics of the proposed methods are analysed in detail.The effectiveness and the efficiency of the proposed methods areproven by the comparison analysis on a number of benchmarknumericalNLPproblems.(3)For the reqirements of model identification and online recedinghorizon optimization in NMPC, the extremal optimization isextendedandfurtherappliedtothesimultaneouslyoptimizationofSVR kernel functions/related parameters and the building ofdymamical model according to multi-step-ahead error. Then,based on the predictive model, a“Horizon based mutation”EO-SQP is designed and served as the online optimizer of aNMPC controller. Finally the simulation on a nonlinear MIMO CSTR is carried out to show the performance of the proposedalgorithm.(4)For a typical nonlinear industrial process - basic oxygen furnaces(BOF) in steelmaking, the proposed neural network model basedon EO-LM learning algorithm is further applied to predict theend-pointproductquality;andtheEO-SQPalgorithmisemployedto find the optimal recipe adjustment. Then an integrated BOFproduction quality control system based on the above mentionedprediction model and optimal recipe adjustment algorithm isdeveloped. The effectiveness of the proposed algorithm is provenbythesimulationresultsonpracticalproductiondata.更多还原