Accurate prediction of superheated steam temperature (SST) is critical for the safe and efficient operation of large-scale thermal power units, particularly under large load variations and high thermal inertia. This study proposes an iTransformer-based SST prediction framework (iTransformer-SST) to address limitations of conventional proportional–integral–derivative (PID) control and existing data-driven models in capturing multivariable coupling, time-delay effects, and physical consistency. Using the A-side subsystem of a 1000 MW thermal power unit, 19-dimensional process data were collected continuously over two months with a sampling interval of 2.4 s. After data preprocessing, time-lagged cross-correlation (TLCC) analysis combined with expert knowledge was employed for feature screening, resulting in ten highly relevant input variables. To enhance predictive robustness, the baseline iTransformer was augmented with a Local Temporal Convolution (LTC) module for local disturbance modeling and a physics-guided regularization term to enforce delayed monotonicity and smoothness constraints. In 240 min rolling forecasts of the final-stage superheater outlet temperature, the proposed model achieved a mean squared error (MSE) of 0.0887, a mean absolute error (MAE) of 0.2312, and a coefficient of determination (R2) of 0.9650, significantly outperforming long short-term memory (LSTM), Informer, and the baseline iTransformer. The combined LTC and physics-guided design reduced MSE by 13.5%, demonstrating strong potential for feedforward-assisted SST control in industrial thermal power applications.
Zhang et al. (Wed,) studied this question.