Quasi-three-dimensional loss prediction model of subsonic compressor cascade based on bidirectional long short-term memory networks and multi-head self-attention

Abstract

The prediction of compressor cascade loss is a crucial aspect of compressor design. Flow separation is an important flow structure and the main source of loss in subsonic cascades. In order to capture the flow separation and accurately evaluate flow loss, a data-driven quasi-three-dimensional (quasi-3D) subsonic compressor cascade loss prediction model based on bidirectional long short-term memory (BiLSTM) and multi-head self-attention is proposed. The model contains four sub-models to predict the pressure, temperature, axial velocity, and total pressure loss coefficient in two-dimensional slices along the axial direction, using Mach number, curved blade angle, solidity, camber angle, and incidence as inputs, respectively. For the purpose of adapting to cascade geometrical change, geometric reformulation is adopted before the model training. The model is trained and tested by validated computational fluid dynamics results, which contain symmetric separation and asymmetric separation samples. It is proved that the model is able to accurately predict flow parameters value in each slice. Then, four typical cases are mainly discussed, which shows that the model can effectively capture the characteristics of flow separation formation and development. Afterward, different models are compared, and it is found that the BiLSTM with multi-head self-attention model achieved the lowest mean squared error, which is because of its outstanding predicting ability in asymmetric separation cases. The work of this paper indicates that the quasi-3D loss prediction model proposed in this paper will be beneficial to the flow separation structure rapid prediction and cascade loss accurate evaluation in compressor design.