A two-stage deep learning method for predicting turbine vane temperature fields under active cooling air flow modulation

Active modulation of turbine cooling air flow is a pivotal technology for enhancing the partial load operational efficiency of gas turbines. However, its practical implementation is impeded by the absence of real-time, precise vane temperature feedback. To address this challenge, this study introduces a novel two-stage deep learning model designed to achieve fast, high-fidelity prediction of turbine vane temperature fields under varying operating conditions and cooling air flow modulation states. The method integrates a gas turbine thermodynamic model and a turbine vane numerical model to generate a comprehensive dataset. The proposed two-stage model first employs a novel upsampling network, termed n-Net, to map low-dimensional inputs representing the gas turbine operating conditions and coolant modulation coefficients to high-resolution temperature fields. Subsequently, an Attention Pix2PixHD network, enhanced with a spatial attention mechanism, performs end-to-end super-resolution to refine the initial predictions and correct local errors. The results demonstrate that the two-stage model achieves excellent predictive performance on the test set, with a maximum relative error of 3.5408 % and a structural similarity index of 0.8460. Crucially, compared to conventional CFD simulations, the model reduces the prediction time by a factor of approximately 105, decreasing the time required per case to around 0.1358 s. The proposed method enables near-real-time, high-fidelity temperature field prediction for applications in active cooling air flow modulation and design optimization.