Current market conditions require flexible gas turbines¹ and especially the ability to operate at low load during times of low electricity prices. Due to the common reduction of firing temperatures with decreasing load, carbon monoxide (CO) emissions limit the achievable engine turndown.² Aging-induced CO emissions may further restrict the market access of a gas turbine, as the CO margin of a given combustion technology to the strict legal limits would be reduced during low load. In order to provide and extend low-load gas turbine operation, reliable and comprehensive gas turbine modeling approaches offer a possibility to understand gas turbine behavior, reduce maintenance costs,³ and enable long-term emission compliance by optimized operation and monitoring. In the case of CO, different modeling approaches exist, such as reactor network modeling, CFD, analytic modeling, or purely statistical modeling, depending on the exact purpose and available computational power and/or data. Statistical modeling can be a powerful and fast approach for data analysis of large data sets. For instance, Turgut⁴ quantified the influence of ambient temperature on CO emissions for different aircraft engines. Similar investigations were previously done by Hung.⁵ Artificial neural networks as a special application of statistical models are also able to predict pollutant emissions when adequate training data are available.⁶ Still, useful information about the governing physical phenomena is not explicitly incorporated in statistical models, which inhibits, for example, their extrapolation capacity.⁷ Though empirical models are commercially available as a replacement for continuous emission monitoring systems (CEMSs). Bainier et al.⁸ recently reported their progress on the development of such a predictive emission monitoring system, which however requires recalibration every quarter of the year. In contrast, physical models such as reactor networks^2,9 and CFD simulations^10–12 are able to predict spatially resolved CO formation and thus allow for detailed physical investigation of combustor design and operation aspects. The accuracy and success of these methods are highly dependent on the existence and quality of boundary conditions, such as oxidant and fuel composition and flow, temperature, pressure, and detailed geometry.¹³ While CFD simulation needs exact component geometry and is computationally expensive, reactor network approaches are usually based on simple geometry and are faster to process, but require additional parameter identification and estimations such as residence times in different reactors, heat loss factors, or unmixedness.¹⁴ Semiempirical analytic CO prediction can deal with a less complete description of the combustion chamber by nature as parameter identification is part of the model setup. However, semiempirical modeling incorporates the underlying physical concepts, while keeping the advantage of very low computational effort. To cite a few examples, Connors et al.¹⁵ developed a semiempirical CO correlation for utility combustion turbines. The correlation features a formulation of characteristic times for CO oxidation and evaporation of the liquid fuel derived from physical laws. The residence time before quenching was, however, assessed by detailed geometry information of the combustion chamber. A Damköhler number, derived by the characteristic times, was shown to correlate with CO emissions by a linear function, which was fit to existing data. Similarly, Lefebvre¹⁶ developed and evaluated a correlation for CO prediction, where flow proportions and operating conditions were implicitly accounted for. The model was successfully tested on different jet engine combustors with varying geometry. More recently, Therkorn et al.¹⁷ assessed burner switch-off concepts for the auto-ignited sequential combustor in the GT24/GT26 gas turbine and successfully modeled its influence on CO emissions with a semiempirical model. Enriching this approach with accurate kinetics in a simple reactor model, Güthe et al.² demonstrated the capability to capture the effect of fuel reactivity (i.e. hydrocarbon and hydrogen content) at lab and engine scale. However, the abovementioned investigations mostly deal with design features and new components. Long-term prediction performance of theoretical models is reported to be an issue, since aging effects may cause deviating operation characteristics.¹⁸ When these aging effects are added to the method, semiempirical models offer an improved long-term predictability. Such models are also suitable for monitoring, since root causes for engine deterioration can be linked online to a real component, as previously shown for NOx emissions.^19,20

中文翻译：

---

重型燃气轮机的长期一氧化碳排放行为：一种基于模型的监视和诊断方法

当前的市场条件要求柔性燃气涡轮机¹，尤其是在电价低时需要在低负载下运行的能力。由于随着负载的降低，燃烧温度通常会降低，因此，一氧化碳（CO）排放会限制可实现的发动机调速。²老化引起的一氧化碳排放可能进一步限制燃气轮机的市场准入，因为在低负荷情况下，将给定燃烧技术的一氧化碳裕度降低到严格的法律限制。为了提供和扩展低负荷燃气轮机的运行，可靠而全面的燃气轮机建模方法为理解燃气轮机的性能，降低维护成本提供了可能，³并通过优化的操作和监控实现长期排放合规性。在CO的情况下，根据确切的目的和可用的计算能力和/或数据，存在不同的建模方法，例如反应堆网络建模，CFD，分析建模或纯粹的统计建模。统计建模可以是对大型数据集进行数据分析的强大而快速的方法。例如，Turgut ⁴量化了环境温度对不同飞机发动机的CO排放的影响。洪以前曾做过类似的调查。⁵当有足够的培训数据时，作为统计模型的特殊应用的人工神经网络也能够预测污染物的排放。⁶尽管如此，有关控制物理现象的有用信息仍未明确纳入统计模型，这会抑制其外推能力。⁷尽管经验模型可从市场上买到，以代替连续排放监测系统（CEMS）。Bainier等。⁸最近报告了他们在开发这种预测性排放监测系统方面的进展，但是，该系统需要每年每季度进行一次重新校准。相反，物理模型，例如反应堆网络^2,9和CFD模拟^10–12能够预测空间分辨的一氧化碳的形成，因此可以对燃烧室的设计和操作方面进行详细的物理研究。这些方法的准确性和成功高度取决于边界条件的存在和质量，这些条件包括氧化剂和燃料的成分以及流量，温度，压力和详细的几何形状。¹³尽管CFD仿真需要精确的部件几何形状并且计算量大，但是反应堆网络方法通常基于简单的几何形状并且处理速度更快，但需要额外的参数识别和估计，例如在不同反应堆中的停留时间，热损失因子或不混合性。¹⁴由于参数识别是模型设置的一部分，因此，半经验分析CO预测可以自然地处理不太完整的燃烧室描述。但是，半经验建模结合了基本的物理概念，同时保留了非常低的计算工作量的优势。举几个例子，康纳斯等。¹⁵开发了用于公用燃气轮机的半经验CO相关性。相关性的特征是根据物理定律制定了CO氧化和液体燃料蒸发的特征时间。但是，淬火之前的停留时间是通过燃烧室的详细几何信息来评估的。由特征时间得出的Damköhler数通过线性函数显示与CO排放量相关，该函数适合现有数据。同样，Lefebvre ¹⁶开发并评估了CO预测的相关性，其中隐含考虑了流量比例和运行条件。该模型已在具有不同几何形状的不同喷气发动机燃烧器上成功测试。最近，Therkorn等人。¹⁷评估了GT24 / GT26燃气轮机中自动点火顺序燃烧器的燃烧器关闭概念，并成功地通过半经验模型模拟了其对CO排放的影响。Güthe等人在简单的反应器模型中用精确的动力学丰富了这种方法。^图2展示了在实验室和发动机规模上捕获燃料反应性（即碳氢化合物含量）影响的能力。但是，上述研究主要涉及设计特征和新组件。据报道，理论模型的长期预测性能是一个问题，因为老化效应可能会导致运行特性出现偏差。^18岁当将这些老化效应添加到方法中时，半经验模型可提供改进的长期可预测性。这样的模型也适用于监视，因为发动机退化的根本原因可以在线链接到实际组件，如先前显示的NOx排放所示。^19,20