燃烧调整是燃气轮机运行的关键技术，确保燃机运行在清洁、稳定的运行区间，其目标是降低燃机的污染物排放，并提高燃机的燃烧稳定性。当环境温度波动时，燃气轮机的运行工况点可能偏离正常范围，导致污染物排放增加，燃烧稳定性降低。因此，每当季节交替时往往需要对升降负荷中的不同工况点进行燃烧优化，以完成燃烧调整。本文针对西门子V94.3A型燃气轮机，利用电厂运行数据建立了燃烧性能参数的预测神经网络。然后基于燃烧性能预测神经网络，利用机器学习优化算法（遗传算法、粒子群算法、模拟退火算法）建立燃烧优化模型，对特定工况点进行优化。燃烧优化的目标参数是NOx排放量、燃烧室振动加速度和燃烧室压力脉动。因此，燃烧优化前必须先掌握目标参数与燃机运行参数之间的映射关系，即建立燃烧性能参数预测模型。传统的基于物理或实验的燃烧性能模型建模复杂，时间成本大且泛化能力一般。故本文基于电厂的运行数据，利用Matlab神经网络工具包建立燃机电站燃烧性能预测神经网络。确定了双层嵌套神经网络的基本结构，确定了输入参数为环境温度、燃气温度、压气机进口导叶开度、预混燃料流量和值班燃料流量，第一层网络的输出参数为机组功率和透平排气温度，第二层输出参数即燃烧优化的目标参数，训练后的模型在较大的工况范围内可以较好地描述燃烧性能。接着通过对燃烧性能参数的敏感性分析和变工况响应分析，研究了输入变量对燃烧性能参数的影响。最后将该模型与机器学习优化算法相结合，开发了燃烧性能优化模型，即通过目标函数和相关约束条件，降低实际工况点地NOx排放，增加燃烧稳定性。模型可以在保证机组功率基本不变和NOx排放不超标的情况下，提高燃烧稳定性，在保证机组功率基本不变和燃烧稳定性参数不恶化的前提下，降低NOx排放。
Combustion adjustment is the key technology of gas turbine operation to ensure that the gas turbine operates in a clean and stable operation range, so the goal of combustion adjustment is to reduce the pollutant emission of gas turbine and improve the combustion stability. When the ambient temperature fluctuates, the operating point of the gas turbine may deviate from the normal range, the pollutant emission will increase, and the combustion stability will decrease. Therefore, it is often necessary to adjust the combustion of the gas turbine whenever the seasons alternate.However, the combustion adjustment technology is firmly grasped by three major gas turbine manufacturers: GE, Siemens and Mitsubishi. Based on years of research and development and operation, each of them has developed its own combustion adjustment system and is developing towards the direction of intelligence. With its monopoly position, the technology is in a state of blockade for our country. Domestic gas turbine power plants can only turn to the technical experts of these companies for each combustion adjustment, which needs to pay a high cost. Therefore, it is very important to develop our own combustion adjustment technology.The target parameters of combustion adjustment are NOx emission, ACC and humming. Therefore, it is necessary to master the mapping relationship (or functional correlation) between target parameters and gas turbine operating parameters before combustion adjustment. The traditional combustion performance model based on physics or experiment is complex, time-consuming. The development of data science provides new ideas for this kind of problems. Based on a large number of operation data of power plant, the combustion performance prediction model of gas turbine station is established by using neural network (ANN). The basic structure of the two-layer nested neural network is determined. The input parameters are ambient temperature, gas temperature, compressor inlet guide vane, premixed fuel flow and pilot fuel flow. The output parameters of the first layer network are unit power (PEL) and turbine exhaust temperature set point (OTC). The output parameters of the second layer network are NOx emission and combustion performance parameters : combustor vibrational acceleration and combustor dynamic pressure. The operation data of F-class power plant is used to train the model, and the trained model can better describe the combustion performance in a large range of operating conditions.Then, through the sensitivity analysis and off design response analysis of combustion performance parameters, the influence of input variables on combustion performance parameters is studied. Finally, a combustion performance optimization model is developed by combining the model with genetic algorithm, that is, through the objective function and related constraints, the NOx emission at the actual working point is reduced and the combustion stability is increased. The model can greatly improve the combustion stability under the condition that the unit power is basically unchanged and NOx emission does not exceed the standard, and reduce NOx emission under the premise that the unit power is basically unchanged and the combustion stability parameters do not deteriorate.
