反问题与神经网络相结合的混流泵叶片优化设计

Blade Optimization of Mixed-Flow Pump Using Inverse Design Method and Neural Network

针对三元扭曲叶片优化设计过程中设计变量较多的问题,采用动量矩为设计变量,再通过反问题计算得到叶片来间接对叶片进行参数化;针对评价函数计算量太大的问题,根据试验设计理论安排训练样本,采用神经网络建立设计变量与目标函数间的复杂的响应关系,并且详细研究了反向传播和径向基函数2种网络在对评价函数进行预测过程中的应用,建立了一种新的叶片优化设计方法.与传统的优化方法相比,其设计变量数目较少,以叶轮内的三维粘性流动分析为基础且大大缩短了计算时间.利用此方法对一台混流泵的扬程和效率进行优化,所得叶片性能良好,从而验证了此方法的有效性.

The ordinary parameterization programs of three-dimensional blade contain a great deal of design variables, therefore an indirect parameterization method was proposed, where the angular momentum was treated as design variables, the blade was calculated by inverse design method, and neural networks were adopted to construct the response relation between the design variable and the objective function. The sample data used to train neural networks were schemed according to design of experiment theory, the application of two kinds of neural networks-back propagation network and radial basic function network were investigated in detail, and a new optimization method was proposed. Compared with the ordinary optimization programs, fewer variables were required in this method based on the three-dimensional viscous computational fluid dynamics(CFD) analysis and the calculation time was shortened obviously. An optimized blade in a mixed-flow pump, where the head and the efficiency were selected as the objective functions, confirms the validity of this newly proposed method.