A Physics-Informed Deep Learning Approach for HDGT Compressor Performance Simulation

Abstract Owing to the complicated structure of a heavy-duty gas turbine, its efficiency is largely impacted by the ambient and operating conditions. Real-time control and predictive maintenance based on performance model becomes a powerful method widely utilized to improve the operating efficiency of the gas turbine. However, the thermodynamic balance-based performance model is hardly applied for real-time control and predictive maintenance of the gas turbine due to its time-consuming computation. In this study, we designed a deep learning operator network (DeepONet) model to simulate the thermodynamic heat balance of axial compressor. The DeepONet is then employed to rapidly predict the efficiency and mass flow of the compressor. First, a physics-based model is built by seamlessly integrating the thermodynamic balance mechanism, structural characteristics, and operating data of the axial compressor. Then, a high-efficiency and high-precision DeepONet model is constructed with reference to the structure of the physics-based model. The yielded surrogate model is utilized to quickly predict key parameters of the compressor including its mass flow and efficiency. The accuracy and effectiveness of the proposed methodology is evaluated by comparing the prediction outputs with measured data through a case study of an axial compressor. A comparison study with conventional long short-term memory model shows that the proposed DeepONet-model provides a promising tool for performance-based smart maintenance of the compressor.