Prediction and optimization of gas turbine secondary air system cooling efficiency based on deep learning

Secondary-air systems (SASs) are critical for maintaining material integrity and optimizing thermal performance in gas turbines (GTs) and related energy equipment. This work introduces an end-to-end framework that couples high-fidelity numerical simulation (NS) with an attention-augmented 1D-CNN (AM-1D-CNN) surrogate and gradient-based optimization to maximize SAS cooling efficiency under realistic bleed-air limits. First, a steady Reynolds averaged Navier–Stokes (RANS) model was validated against extensive experimental data (including LES spot checks at extreme operating points), achieving close agreement between time-averaged Nusselt number predictions and measured values. Next, 632 RANS cases were generated, spanning a wider-than-experimental range of secondary-air mass flows (0.11-2.17 kg/s), inlet temperatures (318.85-500 K), and rotor speeds (Reφ=4.65×105-1.4×106). Two neural architectures (MLP and 1D-CNN) were trained on normalized inputs; the 1D-CNN outperformed the MLP, and embedding a squeeze-and-excitation attention module (AM-1D-CNN) further boosted test-set R2 by 3.65% and reduced RMSE by 31.48%. Permutation-importance (PI) analysis identified secondary-air mass flow, secondary air temperature, and rotor-surface temperature as the dominant predictors. Response-surface modeling then showed that increasing mass flow strongly enhances Nusselt number, while rotor temperature exerts a modest negative influence. To avoid unrealistically large mass-flow solutions, a penalty term was added to the objective, guiding the optimizer toward low secondary-air mass flows that still maximize cooling. Ultimately, optimal boundary conditions were determined within the feasible parameter range. Detailed Computational Fluid Dynamics (CFD) visualizations confirm that the optimized flow not only cools more efficiently but also promotes stable impingement without excessive separation. This framework delivers a rapid, physics-informed pathway to SAS boundary-condition design and establishes a quantitative foundation for future GT cooling-system innovation.Highlights Proposed AM-1D-CNN model improves heat transfer efficiency prediction.Attention mechanism enhances prediction accuracy compared to MLP and 1D-CNN.Response surface and gradient optimization identify optimal boundary conditions.Synergistic effect between key parameters enhances system performance.Optimization process identifies optimal boundary conditions for efficiency.