基于卷积神经网络的燃煤锅炉近壁区H_(2)S浓度分布实时预测模型  

CNN-Based Real-Time Prediction Model for H_(2)S Concentration Distribution near Waterwall of a Coal-Fired Boiler

作　　者：闫靖文 李俊杰 刘欣 李驰 张超群 王赫阳 Yan Jingwen;Li Junjie;Liu Xin;Li Chi;Zhang Chaoqun;Wang Heyang(School of Mechanical Engineering,Tianjin University,Tianjin 300350,China;Yantai Longyuan Power Technology Co.,Ltd.,Yantai 264006,China)

机构地区：[1]天津大学机械工程学院,天津300350 [2]烟台龙源电力技术股份有限公司,烟台264006

出　　处：《天津大学学报（自然科学与工程技术版）》2024年第11期1143-1151,共9页Journal of Tianjin University：Science and Technology

基　　金：国家能源集团科技资助项目(GJNY-21-168).

摘　　要：近年来,燃煤锅炉普遍采用空气分级燃烧技术以降低氮氧化物的排放.空气分级技术的核心是在炉内主燃区形成乏氧的还原性气氛,从而抑制氮氧化物的生成.但还原性气氛会导致强腐蚀性H_(2)S浓度的显著升高,增加了锅炉水冷壁的高温腐蚀风险.由于CFD数值模拟方法耗时较长,目前仍缺少一种能实时、准确地反映锅炉运行过程中近壁区H_(2)S浓度分布的技术手段.针对上述问题,本文首先构建了一个锅炉CFD数值计算模型,对某350MW超临界墙式对冲锅炉近壁区H_(2)S分布特性进行了数值模拟研究,锅炉出口参数及腐蚀位置与现场吻合良好.结果表明,空气分级燃烧技术下炉膛呈还原性气氛,底层对冲燃烧器的对撞气流对侧墙水冷壁的冲刷是造成锅炉侧墙近壁区H_(2)S高浓度的原因.随后,以锅炉的各项运行参数为输入,以近壁区H_(2)S浓度分布图像为输出,构建转置卷积神经网络.基于提出的H_(2)S浓度预测数值模型,搭建了包含120个不同运行工况的数据库,对神经网络进行训练、验证和测试.结果表明,神经网络测试集预测结果与CFD模型预测结果符合良好,30%Local MAPE仅为1.06%,且计算时长在0.1 s以内,实现了燃煤锅炉近壁区H_(2)S浓度分布的实时预测.The air-staged combustion technology has been widely used in coal-fired boilers in recent years to reduce nitrogen oxide emissions.The fundamental principle of air-staged combustion is to inhibit the formation of fuel NOx by creating a reducing atmosphere during the early stages of coal combustion.However,this reducing atmosphere also leads to a significant increase in the concentration of highly corrosive gas H_(2)S,which greatly increases the risk of high-temperature corrosion problems of the boiler waterwall.Despite this issue,currently,methods that can realize the real-time prediction of H_(2)S concentration distribution near the waterwall during boiler operation remain lacking due to the long calculation time of computational fluid dynamics(CFD)simulations.To resolve this problem,a boiler CFD model was first developed to predict the H_(2)S concentration distribution near the waterwall of a 350-MW supercritical wall-fired boiler.The results revealed that the high concentration of H_(2)S near the furnace side walls resulted from the reducing atmosphere created by air-staged combustion and counter-flow air streams from the opposed burners at the bottom level.Subsequently,a deconvolutional neural network that uses the boiler operating parameters as input and H_(2)S concentration distributions near the waterwall as the output was constructed.Utilizing the developed boiler CFD model,a database of 120 operating conditions was constructed for training,testing,and validation of the neural network. The results showed that the H_(2)S concentration distributions predicted by the neuralnetwork model agreed well with those predicted by the CFD model. The 30% Local MAPE was only 1.06%,and thecalculation time was within 0.1 s. This illustrates that the developed neural network model realizes the real-timeprediction of H_(2)S concentration distribution near the waterwall of coal-fired boilers.

关 键 词：高温腐蚀 H_(2)S浓度 计算流体力学 神经网络 

分 类 号：TK229.6[动力工程及工程热物理—动力机械及工程]