基于全连接神经网络的分层旋流火焰燃烧振荡预报

燃气轮机贫油预混低排放燃烧室容易发生燃烧振荡,在宽工况极端条件下燃烧主动控制有潜力成为抑制燃烧振荡的有效方法。然而主动控制系统依赖于对燃烧振荡的快速监测和准确预报,因此有必要针对不同的燃烧振荡预报手段开展研究和验证。本文以甲烷预混同心分层旋流火焰的图像为基础,采用降低图像分辨率和提取火焰结构特征参数这两种不同的方式对火焰图像信息进行简化处理,并使用全连接神经网络对燃烧振荡进行预报研究。结果发现,两种方式都可以较为准确地预报燃烧振荡,精度均达到90%以上。预报精度随着图像分辨率的增加而升高,在极低的图像分辨率(3*3)下,预报精度也能达到90%以上。此外,对根据火焰平均图像提取的结构特征参数进行了敏感性分析,捕捉到了系统稳定性的转变,但参数变化范围受训练集限制。本文提出的基于数据驱动方法对燃烧振荡的预报时间小于2 ms,为实现燃烧振荡实时在线预报提供了支持。

Prediction of Combustion Oscillation Based on Time-Averaged Images of Stratified Swirl Flame Using Fully-Connected Neural Network

Lean premixed low emissions combustion for gas turbine engines is prone to combustion oscillation problems. Active combustion control is likely to effectively suppress combustion oscillation under wide range and extreme operating conditions. However, a successful active control system relies on rapid monitoring and accurate prediction of combustion oscillation. Consequently, relevant research on combustion oscillation prediction based on different methods should be carried out to verify availability. The time-averaged images of stratified swirl flame of CH₄ are adopted here as the basis of our research. In order to simplify the information contented in flame images, two pre-processing methods on images are used here, which are resolution degradation and dimension reduction using flame structure features. The neural network with fully connected layers are implemented to predict combustion oscillation with pre-processed data. It is found that both methods could obtain good prediction accuracy (better than 90%). Besides, the network using degraded resolution as inputs might still behave well with an accuracy over 90% at an extremely low resolution (3*3). A positive correlation of prediction accuracy and resolutions is found. In the terms of flame structure features, the transition of flame stability dynamics are captured under a limited range of parameter variations. The combustion oscillation prediction time is less than 2 ms using the proposed data-driven methods, which provide support for real-time prediction of combustion oscillation.