【摘要】 在压气机运转的实际流动过程中，由于叶片尾迹存在而导致的流量堵塞情况不可忽视。引入非稳态流量系数这一概念，并提出应用RF-GRNN(随机森林-广义回归神经网络)复合代理模型对因尾迹导致的下游流量堵塞情况以及转静叶排间流量进行预测的方法。通过Sobol′敏感性分析法对第一级转子叶片几何参数及转静叶排轴向间距进行了敏感性分析并筛选出敏感参数，采用拉丁超立方抽样法建立起压气机模型数据库。对数据库进行一系列非定常全周数值模拟，并计算得出样本对应的非稳态流量系数。为了实现模型的小样本、高性能预测，引入随机森林(Random Forest)回归算法进行误差校正，训练适用于小样本数据的RF-GRNN复合模型并进行流量预测。结果表明，相比于传统GRNN模型，RF-GRNN复合模型的预测精度获得了显著提升。此复合模型可根据第一级转子叶片最大相对厚度及转静叶排轴向间距两个几何参数来预测下游的流量堵塞情况及转静叶排间流量，在小样本的前提下达到了较高的预测精度，有助于实际工程应用与叶片设计。更多还原

【Abstract】 The mass flow blockage condition caused by wake of blade cannot be neglected in practical engineering.The concept of unsteady mass flow coefficient is introduced, and a hybrid surrogate model: RF-GRNN(Random Forest-Generalized Regression Neural Network) is put out in order to quantitatively predict mass flow of rotor-stator interface and blockage condition downstream caused by wake of rotor.Geometric parametersare screened using Sobol′ Sensitivity Analysis, maximum thickness of rotorblade and axial spacing are chosen as sensitive parameters and database is constructed with Latin Hypercube Sampling method for further simulations.Series of unsteady full annulus CFD simulations were carried out using structured URANS solver.GRNN was used to train and predict database considering its ability to obtain better prediction results when the sample size is small, nevertheless errors between prediction and target still exist.Combined with Random Forest correction model, the RF-GRNN hybrid surrogate model demonstrates high accuracy for predicting unsteady mass flow coefficient with small size of samples.With this hybrid model, we can predict the mass flow of rotor-stator interface and blockage condition downstream only using two geometric parameters: maximum thickness of rotor blade and axial spacing, which contributes to the blade design in practical engineering.更多还原