【作者】 王鹏；

【导师】 司风琪；

【作者基本信息】 东南大学 ， 动力机械及工程， 2023， 博士

【摘要】 大型火电机组的灵活调峰运行,可与间歇性可再生能源发电形成互补,是我国电力行业实现可持续发展和“碳达峰、碳中和”目标的重要努力方向。火电机组的智能监控是保障大型机组长期灵活性运行的有效手段,实时监测设备运行状态,及时发现异常、识别故障并快速制定故障恢复措施,相关研究对于保障火电机组热工过程的安全稳定运行具有重要理论意义与工程应用价值。本文重点围绕基于深度学习的热工过程建模及诊断方法开展研究工作,论文主要研究内容如下:（1）针对机组设备非线性建模过程中数据不平衡问题,提出了一种基于CVAE-GAN网络模型的数据增强方法。首先,基于VAE网络和WGAN-GP网络建立了CVAE-GAN深度生成网络,并引入连续条件因子学习原始数据分布特征。然后,采用增量L₂-偏差计算指标指导CVAE-GAN生成高质量数据,并在数据集过大情况下指导删除冗余数据,从而提高生成数据质量及生成效率。最后,给出了基于数据增强的非线性建模框架,并通过数学仿真过程和火电机组NO_x排放状态监测的案例分析,验证了本文所提数据增强方法的有效性。（2）针对机组设备多模态变工况运行特性,提出了一种融合物理约束的多模态混合建模方法。首先,基于LSTM网络建立了单模态下的数据驱动模型,构建基于物理守恒约束的正则项并引入模型的损失函数中,从而获得单模态下融合物理约束的动态混合模型。然后,采用注意力机制构建了两种多模态切换策略,并对训练后的单模态混合模型进行集成,从而学习电站设备多模态运行特性。最后,给出了基于注意力机制的多模态混合建模框架,并通过火电机组过热蒸汽温度状态监测的案例分析,验证了本文所提建模方法的有效性。（3）针对机组设备异常状态与过程变量和目标变量的相关性,提出了一种基于IBVAE-SR网络模型的非线性系统在线故障检测方法。首先,建立了基于-VAE网络的设备过程参数监测模型,构建了部分隐层变量与目标变量的回归网络,并采用DVIB方法将其在-VAE网络隐层结构中进行扩展,从而得到了融合DVIB理论的IBVAE-SR网络。其次,在网络损失函数中引入可控阈值!和",以保证网络的重构精度、回归效果和特征解耦能力。然后,设计了与IBVAE-SR网络隐层变量和输出残差相对应的特征统计量和残差统计量,用于检测和区分与目标变量相关的故障。最后,给出了基于IBVAE-SR网络模型的故障检测框架,并通过数学仿真过程故障和火电机组磨煤机故障的案例分析,验证了本文所提出的故障检测方法的有效性。（4）针对机组设备运行过程的动态特性,提出了一种基于DIBVAE-SR网络模型的非线性动态系统故障诊断方法。首先,在IBVAE-SR网络基础上,引入基于Bi-GRU的Seq2seq网络结构,从而建立了DIBVAE-SR动态网络模型,并构建基于该网路模型的故障检测框架。其次,在DIBVAE-SR网络中引入空间自注意力网络层,并通过自注意力权重变化获得与故障相关的过程变量,从而实现故障隔离。然后,基于t-SNE降维方法和DBSCAN密度聚类方法,提出了TS-DBSCAN故障分类方法,对DIBVAE-SR网络的隐层特征进行降维处理和进一步的故障分类。最后,给出了基于DIBVAE-SR网络模型的故障诊断框架,并通过CSTH过程故障和火电机组凝汽器故障的案例分析,验证了本文所提出的故障诊断方法的有效性。（5）在现有TPMFDS平台的基础上设计和开发了适用于火电机组实际运行过程的电站设备热工过程建模和故障诊断系统,分别从系统架构、功能交互和现场部署等方面给出了系统构建方案,并给出了基于火电机组SIS数据库的现场开发案例。更多还原

【Abstract】 Flexible peaking operation of large thermal power units can be complementary to intermittent renewable energy,which will be an important effort direction for the electric power industry in our country to realize sustainable development and the goal of "carbon reaching peak,carbon neutrality".Intelligent monitoring of thermal power units is an effective means to ensure the long-term flexible operation of large units.Real-time monitoring of equipment operating status,timely detection of abnormalities,identification of faults,and rapid formulation of fault recovery measures have important theoretical significance and engineering application value for the safety and economy of the entire energy system.This paper focuses on the research about thermal process modeling and diagnosis methods based on deep learning.The main research contents are as follows:(1)Aiming at the data imbalance problem in the nonlinear modeling process of power plant equipment,a data enhancement method based on CVAE-GAN model is proposed.The CVAE-GAN generation model based on VAE and WGAN-GP is established,and the continuous condition is introduced to learn the original data distribution.To improve the quality of generated data and generating efficiency,the enhancement method based on CVAE-GAN model is proposed.The incremental L2-discrepancy can guide the generation training procedure to obtain high-quality generated samples and guide to remove redundant data when the dataset is too large.A numerical simulation case is given to validate the superiority of the proposed model over other common generative models.Then,the proposed model is applied for NOx emission prediction in a coal-fired power plant.(2)Aiming at the multi-modal and variable-condition operation characteristics of power station equipment,a multi-modal hybrid modeling method incorporating physical constraints is proposed.By building an LSTM-based data-driven model in a single modality and introducing constraints based on conservation of physics in the loss function of the model,a dynamic hybrid model incorporating physical constraints in a single modality is constructed.Aiming at more complex multi-modal modeling problems,two multi-modal switching strategies are constructed through the attention mechanism,and the trained single-modal hybrid models are integrated to establish a multi-modal hybrid modeling framework to learn the multimodal operation characteristics of power plant equipment.Taking the temperature monitoring problem of superheated steam in a superheated system as an example,the effectiveness of the proposed modeling method is verified.(3)Aiming at the correlation between the abnormal state of power plant equipment and process variables and target variables,an online fault detection method for nonlinear systems based on IBVAE-SR model is proposed.Considering the extensibility of the hidden layer structure of β-VAE network,the IBVAE-SR model integrating the deep variational information bottleneck theory is proposed.To improve the efficiency of model training,the equilibrium coefficient β in the model loss function is fixed and the threshold C to be trained is introduced to ensure the reconstruction accuracy and feature composition ability of the model.To improve the effect of fault detection,the feature statistics and residual statistics corresponding to hidden layer variables and model residuals are designed to detect and distinguish faults related to or unrelated to the target variables.Taking mathematical simulation as a verification example and coal mill system as an application example,the validity of the fault detection method based on the IBVAE-SR model proposed in this paper is verified.(4)Aiming at the spatiotemporal characteristics of thermal power units,a fault diagnosis method for nonlinear dynamic systems based on DIBVAE-SR model is proposed.Based on the IBVAE-SR model proposed in Chapter 4,the dynamic model of DIBVAE-SR is established by introducing the Seq2 seq network structure based on Bi-GRU,and the fault detection framework based on this model is constructed.To achieve fault isolation,a spatial self-attention network layer is introduced in the model,and the process variables related to faults are judged by observing the changes of self-attention weights.By using the t-SNE method and the DBSCAN density clustering method,feature visualization and fault classification are performed on the hidden layer variables of the model.Taking the CSTH process as the verification example and the cold end system as the application example,the effectiveness of the fault diagnosis method based on the DIBVAE-SR model proposed in this paper is verified.(5)A thermal process modeling and fault diagnosis system that can be applied in the field operation process of thermal power units is designed and developed.The system construction scheme is given from the aspects of system architecture,functional interaction and field deployment,and a field development case based on the SIS database of thermal power units is given.更多还原