Fast and accurate prediction of flow in multi-row cascade based on combined neural networks

This paper proposes a fast and accurate method for predicting multi-row cascade flow based on a framework of combined neural networks. The primary idea of this method is to decompose the whole-annulus of multi-row cascade into different types of sub-regions, and flow prediction surrogate models based on neural networks are constructed for these sub-regions in a rectangular computational domain by applying the coordinate transformation technique. The prediction surrogate models for each sub-region are then combined, and the continuity of flow at the interfaces among sub-regions is used to iteratively compute the whole-annulus flow in the multi-row cascade. The main advantages of the proposed method include reduced dataset generation cost and neural network training cost through spatial decomposition, as well as the ability to achieve fast prediction of whole-annulus flow in multi-row cascade by combining the neural network surrogate models of the sub-regions. The test case of a two-dimensional stator-rotor interaction indicates that the prediction time of the developed method is approximately 5% of that required for numerical simulation, with over 99% of the nodes in the flow field exhibiting a normalized absolute error of less than 0.05. This approach can be further extended to the fast prediction of three-dimensional flow in multi-stage turbomachinery.