基于在线支持向量机的锅炉燃烧系统动态建模

论文总字数：38457字

摘 要

在当今社会高能源需求和高环保要求的重压下，在以煤炭为主的中国能源结构难以改变的背景下，作为能源产出和煤炭消耗单位，燃煤电厂面临迫切的节能减排形势。锅炉燃烧优化技术作为电厂提高燃料能量利用效率、降低污染物排放量的重要途径，愈发受到重视。基于数据驱动的数学建模成本低，无需求解复杂的微分方程，算法稳定性较好，在锅炉燃烧优化技术的实践中具有非常大的优势。

锅炉燃烧系统涉及复杂的传热传质过程，呈现强非线性，与此同时受到众多外部因素的影响，难以通过燃烧机理建立动态模型。然而燃烧系统的输入与输出数据一定程度上反映了实际系统的动态特性，为数据建模方法提供了现实依据。锅炉燃烧模型是锅炉燃烧优化的基础，现有方法建立的模型大多为离线模型、稳态模型，对于系统工况变动的适应性较差，难以有效地呈现锅炉燃烧系统的动态特性，因此，研究满足锅炉燃烧优化要求的在线动态建模方法十分必要。支持向量机建模是一种以结构风险最小化为基本原则的非线性模型辨识方法，具有泛化性好、不易陷入局部极小点等优点，在电厂热工过程系统建模领域得到了成功的应用。支持向量机本身也不断发展，近年来学者们创造性地提出了准确在线支持向量回归（AOSVR）建模方法，在非线性自适应建模领域得到了广泛的关注，利用该方法可实现锅炉燃烧系统的非线性模型的实时在线更新。

本文研究基于在线支持向量机的锅炉燃烧系统动态建模，主要的研究内容包括：

1. 采用增量式学习方法建立锅炉效率和排放量“多输入—单输出”的支持向量机模型，并进行模型参数实时在线更新，同时提出一种离散选取训练集样本的方法。
2. 比较批量式和增量式2种学习方法的差异，对比离线模型和在线模型的优劣，分析在线模型训练集先进先出和子集先进先出2类样本剔除规则的区别。结果显示，对锅炉燃烧系统使用增量式学习方法建立遵循子集先进先出规则的在线模型效果最好。
3. 采用某电厂600MW四角切圆锅炉燃烧系统的运行和试验数据及MATLAB计算机仿真，验证模型效果。结果表明，在线支持向量机动态模型精度符合要求，具有对于复杂工况的自适应性和在线建模算法实现的时效性。

关键词：锅炉燃烧建模；动态建模；支持向量机；批量式学习；增量式学习

Abstract

In modern society, under the pressure of high power demand and district environment-protection criteria, facing the fact that Chinese coal-dominant power structure is hard to change, energy-saving and emission-reducing have been increasingly urgent for coal-fired power plants, which have massive energy production and require large coal consumption. As the significant way for power plants to improve power efficiency and reduce pollution emission, attention people pay on boiler combustion optimization technology is rising. Data-driven modeling optimization requires low cost, no need for solving complex differential equations, and has excellent stability hence it has huge advantages upon boiler combustion optimization technologies in practice.

Boiler combustion involving complicated heat and mass transfer processes, presents to be highly non-linear. Meanwhile, under the influence of many external contributory factors, it is hard to modeling boiler combustion by combustion mechanism. However, inputs and outputs of combustion system partly reflect the dynamic characteristics of actual system, which provides the practical basis for data-driven modeling method. Boiler combustion modeling is the basis of boiler combustion optimization. Most models by present technologies are offline or static, which can hardly reflect the dynamic process effectively with poor adaptability for system variations in working conditions. Therefore, it’s necessary to research on a method for online and dynamic modeling catering to the requirement of boiler combustion optimization. Support Vector Machine (SVM) is a non-linear model identification approach based on Structure Risk Minimization (SRM), has extraordinary generalization ability and can avoid being trapped into local minima, hence it is successfully used in thermal engineering processes modeling of power utilities. SVM itself has been developing. In recent years, scholars have creatively put forward the Accurate Online Support Vector Regression (AOSVR), which has been widely concerned in non-linear adaptive modeling. Non-linear model of boiler combustion system can be updated real time by AOSVR.

This paper is about dynamic modeling of boiler combustion system based on online support vector machine. The main research contents include:

1. Incremental learning is adopted to establish Multi-Inputs Single-output (MISO) SVM models of boiler efficiency and emission, and update model parameters real time. A method is raised to select training samples discretely.
2. Comparing the difference between batch and incremental implantation, and compares the pros and cons between FIFO and Sub-FIFO of online models. The result of on-line modeling boiler combustion system by using incremental learning and Sub-FIFO turns out to be the best.
3. Using the operational and experimental data of some power plant’s 600MW tangential firing boiler combustion system and computer simulation by MATLAB, the model can be justified. It turns out to be that Online Support Vector Machine (OSVM) equips the dynamic model with qualified accuracy, appreciable adaptability for complex variations in operating conditions and efficiency of on-line modeling algorithm.

Key words: Boiler Combustion Modeling；Dynamic Modeling；Support Vector Machine；Batch Learning；Incremental Learning

目录

摘要 I

Abstract II

目录 IV

第一章 绪论 1

1.1 研究背景及意义 1

1.2 锅炉燃烧建模研究现状 2

1.2.1 燃烧系统的机理建模 2

1.2.2 燃烧系统的辨识建模 2

1.2.3 燃烧系统的在线建模 3

1.3 在线支持向量机的发展和应用 3

1.4 本文研究的主要内容 5

第二章 支持向量机理论 6

2.1 支持向量分类机 6

2.1.1 线性支持向量分类机 6

2.1.2 非线性支持向量分类机 7

2.2 支持向量回归机 7

2.2.1 线性支持向量回归机 8

2.2.2 核函数 9

2.2.3 非线性支持向量回归机 10

2.2.4 支持向量回归机求解方法 11

2.3 准确在线支持向量回归 12

2.4 动态建模方法 17

第三章 锅炉燃烧系统建模 18

3.1 锅炉燃烧系统 18

3.1.1 总体情况 18

3.1.2 燃烧器及二次风门布置 18

3.2 模型数据 20

3.2.1 模型数据组成 20

3.2.2 模型数据预处理 20

3.3 燃烧模型输入输出 20

3.4 建模参数选取 22

3.4.1 网格法和折交叉验证法 22

3.4.2 确定建模参数 23

3.5 离线模型建立思想 24

3.6 在线模型建立思想 24

第四章 程序实现与计算机仿真 26

4.1 建模程序具体实现 26

4.1.1 批量式学习算法 26

剩余内容已隐藏，请支付后下载全文，论文总字数：38457字