航空学报 > 2023, Vol. 44 Issue (5): 126816-126816   doi: 10.7527/S1000-6893.2022.26816

神经网络宽度对燃烧室排放预测的影响

王志凯1,2(), 陈盛1, 范玮2   

  1. 1.中国航发湖南动力机械研究所,株洲 412002
    2.西北工业大学 动力与能源学院,西安 710129
  • 收稿日期:2021-12-14 修回日期:2021-12-30 接受日期:2022-01-10 出版日期:2022-01-20 发布日期:2022-01-18
  • 通讯作者: 王志凯 E-mail:nhwzk12@126.com
  • 基金资助:
    国家级项目

Effect of neural network width on combustor emission prediction

Zhikai WANG1,2(), Sheng CHEN1, Wei FAN2   

  1. 1.AECC Hunan Aviation Powerplant Research Institute,Zhuzhou 412002,China
    2.School of Power and Energy,Northwestern Polytechnical University,Xi’an 710129,China
  • Received:2021-12-14 Revised:2021-12-30 Accepted:2022-01-10 Online:2022-01-20 Published:2022-01-18
  • Contact: Zhikai WANG E-mail:nhwzk12@126.com
  • Supported by:
    National Level Project

摘要:

为研究神经网络宽度对航空发动机燃烧室排放预测的影响,基于录取的全环燃烧室排放试验数据,构建了以燃烧室进口空气温度、进口空气压力、供油流量和油气比为输入,CO和NO x 排放指数为输出的神经网络预测模型,确定了最优网络宽度。结果表明,存在一个最优的网络宽度值,使得神经网络预测模型的拟合优度和预测精度达到最佳,本文最优的网络宽度为24。通过拟合优度、误差分析和敏感性分析验证了构建的神经网络模型的准确性和泛化性。基于最优网络宽度的神经网络预测模型能够很好地挖掘输入参数与排放指数之间的映射关系,可作为给定工况参数下燃烧室排放预测工具。最后,基于敏感性分析,结合燃烧物理机理和试验现象对构建的神经网络可解释性进行了探讨。

关键词: 燃烧室, 排放, 神经网络宽度, 预测模型, 可解释性, 敏感性分析

Abstract:

To study the effect of neural network width on aero-engine combustor emission prediction, a neural network prediction model is established based on the experimental data of full annular combustor emission. The inlet air temperature, inlet air pressure, fuel flow rate and fuel-air ratio are defined as the input parameters, and CO and NO x emission index as the output parameters of the model. Results indicate that there is an optimal network width value, which makes the best goodness of fit and prediction accuracy of the neural network prediction model, and the optimal network width in this study is 24. The accuracy and generalization of the neural network model are verified by goodness of fit, error analysis and sensitivity analysis. The neural network prediction model based on the optimal network width can well mine the mapping relationship between input parameters and emission index, and can be used as a prediction tool for combustor emission with given operating parameters. Finally, based on sensitivity analysis, the interpretability of the constructed neural network is discussed in combination with the combustion physical mechanism and experimental phenomena.

Key words: combustor, emission, neural network width, prediction model, interpretability, sensitivity analysis

中图分类号: