Fast Prediction Model for Gas Turbine Metal Temperature in Digital Twin Applications

To address the real-time prediction requirements of gas turbine metal temperature in digital twin applications, a fast prediction method of gas turbine metal temperature based on recurrent neural networks (RNN) was proposed. Conventional physics-based whole-engine temperature field models driven by operational curves suffer from high complexity and long computation time, failing to meet the demands of real-time prediction and fault warning in gas turbine testing and operation. Two models were trained using datasets from physics-model simulations and whole-engine experiments, achieving fidelity of 99.14% and 94.32% after optimization. The influence of neural network hyperparameters on model performance was systematically analyzed, and effective tuning methods for model optimization and overfitting reduction were proposed. Model generalization capability was verified through test datasets. Deployment was implemented in a prototype F-class heavy-duty gas turbine using functional mock-up interface (FMI) standards, providing an effective solution for real-time temperature prediction in gas turbine digital twin systems.