摘要
为克服计算流体力学(CFD)方法计算成本高及无法重复利用计算结果的缺陷,基于深度学习方法,利用132组二维流场数据建立NACA0018翼型α=2°~8°、Re=0.1×106~1.6×106下压力场、速度场的神经网络定常预测模型。以此为基础,将低速不可压缩流动能量守恒方程作为约束条件,考虑到翼型升阻力与表面压力的相关性及流场静压与动压的关系,提出一种能间接约束翼型压力场与速度场之间关系的激活函数。结果表明,传统激活函数与改进激活函数下的神经网络在压力场预测平均误差均在2.77%左右,但传统神经网络在速度场预测中平均误差达到11%,最大达到26.993%,而改进激活函数平均误差只有2.77%。与传统的激活函数相比,改进的激活函数的神经网络因存在内部间接约束,故在翼型速度场的预测中更准确,流场过渡更均匀,并且神经网络方法可以通过保存模型以达到重复利用数值模拟结果目的,相比传统CFD方法数小时的计算,训练完成的神经网络只需要数秒即可得到计算结果,可大大减少计算时间。
In order to overcome the disadvantages of the computational fluid dynamics(CFD)method such as high computational cost and the inability to reuse the computational results,a steady prediction model of pressure and velocity fields for NACA0018 airfoil inα=2°-8°,Re=0.1×106-1.6×106 is established based on deep-learning method using 132 sets of two-dimensional flow data.The energy conservation equation of incompressible flow at low velocity is used as the constraint condition.Considering the correlation between lift drag and surface pressure,an activation function is proposed.The results show that for pressure field prediction the average error of the traditional neural network is about 2.77%,but for velocity field prediction that of the traditional neural network is 11%and the maximum is 26.993%,while the average error of the improved neural network is only 2.77%.Compared with the traditional activation function,the improved activation function neural network is more accurate in predicting airfoil velocity field and the flow field transition is more uniform.Compared with the traditional CFD method,the neural network can obtain the flow field in a few seconds,which can greatly reduce the calculation time.
作者
杨从新
凌祖光
王岩
钱晨
赵斌
周楠楠
YANG Congxin;LING Zuguang;WANG Yan;QIAN Chen;ZHAO Bin;ZHOU Nannan(College of Energy and Power Engineering, Lanzhou University of Technology, Lanzhou 730050, China;Nuclear Power Institute of China, Chengdu 610000, China)
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2021年第3期20-28,共9页
Journal of Xi'an Jiaotong University
基金
甘肃省风力机研发专项基金资助项目(071904)。
关键词
神经网络
深度学习
流场预测
多任务回归
neural network
deep learning
flow field prediction
multi task regression