Enhancing rotor temperature field prediction for steam turbines using EC-SSC-DNN: A structural-symmetry-constrained deep learning framework

Accurate and timely prediction of rotor temperature field is essential for the stable and efficient operation of steam turbines. To address the lack of physical interpretability in conventional models, EC-SSC-DNN, a novel deep learning framework, is proposed, incorporating structural symmetry constraints and an error correction mechanism. A multi-source dataset of 90 operating conditions is constructed by combining Distributed Control System (DCS) data with Computational Fluid Dynamics (CFD) simulations. The conditions are classified into five categories based on load variation during startup, with representative samples extracted using a Gaussian Mixture Model (GMM). A structural-symmetry-constrained Deep Neural Network is then developed, incorporating a physics-informed loss function and an error correction module to address prediction deviations caused by multi-parameter coupling. Experimental results show that EC-SSC-DNN achieves high accuracy in predicting the temperature field during turbine startup, with mean absolute errors (MAE) ranging from 7.5% to 15.6% and R2 values from 0.96 to 0.99 across representative conditions. This framework provides both theoretical insights and practical support for ensuring the stable operation of steam turbines.