电是一种至关重要的战略性资源，关乎国民经济命脉、国家能源安全。火力发电作为主体，提质增效减少污染是供给侧结构性改革的重要举措。而烟气含氧量是反映风煤比的重要参数，是锅炉热效率计算和系统优化的重要指标，同时也是确保锅炉系统经济和安全的前提。但烟气含氧量的测量存在成本高、过程复杂、传感器易损坏、及时性低和精度日益下降等问题。本文以陕西榆林某燃煤电厂为背景，结合锅炉工艺，构建数据库并对机器学习和深度学习的软测量建模方法进行深入研究。具体研究内容如下：

1）通过分析烟气含氧量的化学原理和锅炉工艺并进行实地考察，将烟气含氧量作为主导变量初步选取合理的辅助变量，同时采集燃煤电厂200MW的亚临界自然循环锅炉1号机组的实际生产数据。对所有样本数据进行预处理，同时引入灰色关联分析(grey relational analysis, GRA)模型进行辅助变量的精选，并划分数据集构建模型数据库。

2）针对传统的机器学习建模方法，首先，建立支持向量机(support vector machine, SVM)模型；其次，针对SVM参数优化问题，融合了粒子群算法(particle swarm optimization, PSO)，并加入了自适应权重、异步学习和压缩因子对PSO进行改进，进而嵌入到SVM中构建改进PSO-SVM软测量模型对惩罚因子与核函数参数进行寻优；最后，结果表明基于改进PSO-SVM模型的可靠性较好、精度高，能更加快速地寻找到参数优化的组合值，并有效避免粒子群陷入局部最优解问题或者陷入停滞的问题。

3）针对深度学习的建模方法，提出一种基于长短时记忆网络(long short-term memory, LSTM)的软测量模型。首先，为解决网络模型优化问题，提出一种超参数联合寻优的策略，对神经元个数、时间步幅和网络深度进行全局和局部寻优，构建了性能更优的改进PSO-LSTM软测量模型；其次，仿真实验及对比分析表明该模型较改进PSO-SVM模型，精度更高、寻优简便且泛化性性能更好；最后，将该模型融入到燃煤电厂实际应用中，基于PyQt5设计了基于LSTM的烟气含氧量软测量系统。

烟气含氧量软测量的方法研究具有重要的现实意义，本文所提出的方法满足燃煤电厂烟气含氧量高精度监测的需求，可作为代替氧化锆传感器的有效方法。

Electricity is a vital strategic resource, which is related to the lifeline of national economy and national energy security. As the main body of thermal power generation, improving quality and efficiency to reduce pollution is an important measure of supply-side structural reform. The oxygen content of flue gas is an important parameter reflecting the ratio of air to coal, an important index for boiler thermal efficiency calculation and system optimization, and a prerequisite for ensuring the economy and safety of boiler system. However, the measurement of oxygen content in flue gas has many problems, such as high cost, complex process, easy damage to sensor, low timeliness and decreasing precision. Based on a coal-fired power plant in Yulin, Shaanxi province, this paper builds a database based on boiler technology and makes an in-depth study on soft measurement modeling methods of machine learning and deep learning. The specific research content is as follows:

1) The chemical principle of oxygen content in flue gas and boiler technology were analyzed and field investigation was carried out. The oxygen content of flue gas was taken as the dominant variable and a reasonable auxiliary variable was selected. At the same time, the actual production data of unit 1 of subcritical natural circulation boiler of 200MW power plant were collected. All the sample data were preprocessed, and the grey relational analysis (GRA) model was introduced to select the auxiliary variables, and the model database was constructed by dividing the data sets.

2) Aiming at the traditional machine learning modeling method, the support vector machine (SVM) model is firstly established. Secondly, for the SVM parameter optimization problem, the particle swarm optimization (PSO) is fused and the adaptive weight, asynchronous learning and compression factor are added to improve the PSO. Then embedded in the SVM to construct improved PSO-SVM soft sensor model to optimize the penalty factor and kernel function parameters. Finally, the results show that the improved PSO-SVM model has good reliability and high precision, can find the combination value of parameter optimization more quickly. And can effectively avoid the problem of particle swarm falling into local optimal solution or falling into stagnation.

3) Aiming at the modeling method of deep learning, a soft sensor model based on long short-term memory (LSTM) is proposed. Firstly, to solve the problem of network model optimization, a hyper-parameter joint optimization strategy was proposed. Based on global and local optimization of units, time step and LSTM layers, an improved PSO-LSTM soft sensor model with better performance was constructed. Secondly, simulation experiments and comparative analysis show that this model has higher accuracy, simple optimization and better generalization performance than the improved PSO-SVM model. Finally, the model is integrated into the practical application of coal-fired power plant. Based on PyQt5 design, a soft sensor system based on LSTM is developed.

The method of soft sensor of oxygen content in flue gas has important practical significance. The method presented in this paper meets the demand of high precision monitoring of oxygen content in flue gas of coal-fired power plants, and can be used as an effective method to replace the zirconia sensor.