[Objective

]

Coal-fired power units typically exhibit strong nonlinearity

long time delays

and pronounced physical couplings during operation. These characteristics make it difficult for traditional mechanismbased or purely data-driven modeling approaches to simultaneously achieve high dynamic accuracy

strong generalization capability

and low deployment cost in real industrial environments. To address these challenges

this study proposes a physics-informed dual-branch convolutional transformer network(PID-CTNet) that integrates physical priors with deep learning to achieve collaborati

ve prediction of multiple combustion-related variables. [Methods

]

First

a multi-source feature space is reconstructed based on combustion and thermodynamic mechanisms. Twenty-seven key physical variables are selected from four categories(load commands

fuel and airgas system parameters

thermodynamic properties of the working fluid

and historical feedback indicators) to form a physically meaningful input representation. Second

a multi-task deep learning framework combining convolutional neural networks(CNN) and Transformer is developed. The front-end temporal convolutional networks(TCN) extract local transient features and perform dimensionality reduction

while the back-end Transformer introduces a self-attention mechanism with a global receptive field to capture long-term dependencies and large-delay characteristics inherent in boiler dynamics. Finally

a hard parameter-sharing strategy is employed to enable simultaneous prediction of NO

x

SO

2

and CO emissions

as well as main and reheat steam temperatures

within a unified modeling framework. [Results

]

Experimental results demonstrate that the proposed end-to-end model achieves high prediction accuracy

with root mean square error of 2.895

mean absolute error of 1.654

mean absolute percentage error of 0.565%

and coefficient of determination R

2

of 0.957 on the test set. The prediction errors of all five industrial key indicators remained below 3%

significantly outperforming traditional regression and time-series models such as SVR and LSTM

as well as single-task deep learning baselines. Ablation studies further validate the effectiveness of incorporating physical priors

the complementary roles of the TCN-Transformer hybrid architecture

and the performance gains brought by the multi-task learning mechanism. These components jointly enhance model accuracy

stability

and robustness under complex operating conditions

including load fluctuations

fuel variations

and frequent operational adjustments. [Conclusion

]

The PID

-CTNet model effectively integrates physical mechanisms with deep learning

achieving a favorable balance between prediction accuracy and computational efficiency. By providing lightweight and reliable collaborative prediction of emissions and steam temperature parameters

the proposed method offers a practical and deployable solution for intelligent operation

emission optimization

and safety control of coal-fired power units. The results demonstrate that physics-informed multi-task learning is a promising direction for improving the interpretability

engineering applicability

and longterm maintainability of data-driven models in modern power plant operation and optimization.