轴流压气机通流计算中代理模型的应用

Application of Surrogate Models for Through-Flow Calculation in an Axial-Flow Compressor

关键词:  代理模型  通流计算  轴流压气机  经验模型  支持向量机  高斯过程

Key words:  Surrogate model  Through-flow calculation  Axial-flow compressor  Empirical model  Support vector machine  Gaussian process

摘要:

在轴流压气机初始设计阶段，二维通流计算方法可以快速评估压气机的性能，其准确性高度依赖于经验模型的预测精度。当压气机超出经验模型适用范围时会增大模型的预测误差。为提高预测精度，研究了在通流计算中采用代理模型取代传统经验模型的可行性。以双级跨声轴流速压气机的几何参数和气动实验数据作为代理模型的训练样本库，采用敏感性分析方法确定最重要的特征参数作为输入。分别基于支持向量机回归（SVR）和高斯过程回归（GPR）两种监督学习方法构建代理模型，并在模型训练中通过贝叶斯优化算法寻找模型最佳超参组合。将代理模型集成到基于流线曲率法（SLC）的通流程序中，对该压气机特性进行评估，并与传统经验模型计算结果相对比。对比显示，相较于传统经验模型，SVR和GPR两种代理模型分别降低了62.2%与55.2%的总压特性平均预测误差以及48.4%与50.1%的绝热效率特性平均预测误差。对比结果表明，当超出传统经验模型适用范围时，代理模型不失为一种可靠的替代方案。

Abstract:

The two dimensional through-flow method can predict the performance of the compressor fast during the preliminary design period. Its accuracy is highly dependent on the empirical models. The prediction error increases if the compressor is far beyond the applicable range of empirical models. In order to improve the prediction accuracy of though-flow method， the feasibility of using surrogate model to replace the traditional empirical was studied. The geometric parameters and aerodynamic experimental data of a two-stage transonic compressor were used as the training database for surrogate models. The most influential features were selected as inputs using sensitivity analysis method. Two supervised machine learning methods， support vector machine regression （SVR） and Gaussian process regression （GPR）， respectively， were implemented to build the surrogate models. Bayesian optimization algorithm was applied to search the optimal hyper parameters. The trained surrogate models were integrated into the through–flow program based on the streamline curvature method （SLC） to evaluate the characteristics of the compressor and compare it with the calculation results of the traditional empirical models. The comparison showed that， compared with the traditional empirical model， the SVR and GPR surrogate models reduced the average prediction error of the total pressure characteristics by 62.2% and 55.2%， and the adiabatic efficiency characteristics by 48.4% and 50.1%， respectively. The results indicated that the surrogate model is a reliable alternative when the compressor works beyond the applicable range of traditional empirical models.