航空学报 > 2021, Vol. 42 Issue (4): 524331-524331   doi: 10.7527/S1000-6893.2020.24331

基于深度学习的单排孔气膜冷却性能预测

李左飙, 温风波, 唐晓雷, 苏良俊, 王松涛   

  1. 哈尔滨工业大学 能源科学与工程学院, 哈尔滨 150001
  • 收稿日期:2020-06-01 修回日期:2020-06-30 发布日期:2020-07-27
  • 通讯作者: 温风波 E-mail:fengbo@hit.edu.cn
  • 基金资助:
    国家科技重大专项(2017-I-0005-0006)

Prediction of single-row hole film cooling performance based on deep learning

LI Zuobiao, WEN Fengbo, TANG Xiaolei, SU Liangjun, WANG Songtao   

  1. School of Energy Science and Engineering, Harbin Institute of Technology, Harbin 150001, China
  • Received:2020-06-01 Revised:2020-06-30 Published:2020-07-27
  • Supported by:
    National Science and Technology Major Project(2017-I-0005-0006)

摘要: 气膜冷却是增强涡轮叶片的高温耐受力,间接提高涡轮进口温度的有效手段之一。目前气膜冷却孔布局的主流设计方法是先通过计算流体力学(CFD)筛选和优化初始方案,再进行模型实验。这种方法设计周期长,时间成本高。传统上用于快速评估冷却效率的经验公式法存在函数形式复杂,拟合精度有限,参数适用范围较窄等问题。因此基于深度学习原理,设计了一种基于多层感知器模型(MLP)的深度神经网络,建立了绝热气膜冷却效率的预测模型。使用CFD数据训练网络,结果表明:深度学习模型在训练集和验证集上具有大于0.95的拟合度,在测试集上具有大于0.99的拟合度,可以较好地识别数据集中的抽象特征,具有较高的精度和较好的泛化能力。此外,在满足精度要求的前提下,一个完成训练的深度学习模型能够有效减少预测耗时,提高预测效率,在快速评估冷却布局性能方面具有较好的应用前景。

关键词: 气膜冷却, 绝热气膜冷却效率, 计算流体力学(CFD), 神经网络, 深度学习, 多层感知器(MLP), 预测

Abstract: Film cooling is an effective way to enhance the high temperature resistance of turbine blades and indirectly increase the inlet temperature of the turbine. Currently, the mainstream design method of the film cooling hole layout is to select and optimize the initial schemes with Computational Fluid Dynamics (CFD), followed by model experiments. However, this method has a long design period and thus is time-consuming. The traditional empirical formula method for rapid evaluation of cooling efficiency is problematic in that the function form is complex, the fitting precision is low and the applicable range of parameters is narrow. This paper designs a deep neural network with a Multi-Layer Perceptron model (MLP) based on the principle of deep learning, and establishes a prediction model for adiabatic film cooling efficiency. The calculation of CFD is utilized to train the network, and the results obtained from the modeling efforts indicate that the deep learning model has a fitting degree of more than 0.95 on the training set and the verification set, and more than 0.99 on the test set. Hence, it can effectively identify the abstract features of the data set with high precision and good generalization ability. Additionally, on the premise of meeting the requirement of sufficient precision, the deep learning model takes only 1/1000 of the time of CFD to complete the prediction process, exhibiting a good application prospect in the rapid evaluation of cooling layout performance.

Key words: film cooling, adiabatic film cooling efficiency, Computational Fluid Dynamics (CFD), neural networks, deep learning, Multi-Layer Perceptron (MLP), prediction

中图分类号: