基于人工神经网络的离心泵叶轮边界涡量流预测

Boundary vorticity flux prediction of the centrifugal pump impeller based on artificial neural network

根据边界涡量动力学理论,从边界涡量流在离心泵叶轮内表面的分布情况,可获知叶轮的受力状况,进而改进叶轮设计.以BP神经网络和径向基神经网络为建模手段,以叶轮内表面的边界涡量流为预测目标,通过高精度的CFD计算获得70个离心泵叶轮内表面的BVF分布,建立可用于训练人工神经网络的初始样本集;再利用63个初始样本建立离心泵叶轮几何参数和边界涡量流的非线性映射关系,并用剩余的7个校对样本进行测试.根据神经网络预测结果和数值模拟计算结果的误差分析,确定最适用于离心泵叶轮边界涡量流预测的神经网络类型.研究表明:径向基(RBF)神经网络的预测精度高于BP神经网络,其训练时间更短、运行稳定性更高;径向基函数的宽度对RBF神经网络的预测性能有较大影响,当径向基函数宽度取0.3时,RBF神经网络的预测性能最佳,预测误差仅0.0203;RBF神经网络预测所得叶轮内表面的边界涡量流分布,可以作为评价叶轮水力设计优劣的重要指标,进而指导叶轮机械的优化设计.

Based on the theory of boundary vorticity dynamics,the force and moment of an arbitrary shape object can be obtained by the boundary vorticity flux integral.Therefore,if the distribution of the boundary vorticity flux on the inner surface of the centrifugal pump impeller can be predicted,the internal flow condition and force of the impeller can be analyzed and known,which helps to improve the design of the impeller.Hence,the BP neural network and radial basis neural network were chosen as the modeling methods,and the boundary vorticity flux on the inner surface of impeller was taken as the prediction target.Firstly,high-accuracy CFD calculations were performed to obtain the BVF distributions in 70 centrifugal pump impellers and construct 70 groups of initial training samples.Then,63 initial samples were used to establish the nonlinear mapping relation between the geometric parameters of the centrifugal pump impeller and the boundary vorticity flux.Besides,the remaining 7 proofread samples were used to test the relation by comparing the predicted values of the neural networks with the calculated values of numerical simulations.According to the magnitude of the error,the predictive performance of the artificial neural network was evaluated.It shows that compared with BP neural network,RBF neural network has higher prediction accuracy,shorter training time and higher opera-tion stability.The width of radial basis function has a great influence on the prediction performance of RBF neural network.When the radial basis function width is set as 0.3,the RBF neural network has the best prediction performance and the prediction error is only 0.0203.The boundary vorticity flux distribution on the inner surface of the impeller predicted by the RBF neural network can be used as an important index to evaluate the hydraulic design of the impeller,and then guide the optimal design of the turbomachinery.