[Objective

]

High-precision prediction of sulfur dioxide(SO

2

) mass concentration in flue gas at outlet of the wet flue gas desulfurization(WFGD) systems is a fundamental prerequisite for achieving both environmental compliance and energy-saving optimization control in modern thermal power plants. Currently

the industry relies heavily on traditional data-driven black-box prediction models. However

these models exhibit significant limitations

particularly poor physical interpretability and inadequate generalization performance under complex and fluctuating variable operating conditions. Due to their reliance on purely statistical correlations

black-box architectures often capture spurious mappings rather t

han underlying causalities

leading to predictions that may violate basic chemical kinetics and mass transfer principles. [Methods

]

To address these challenges

this study proposes a novel physics-informed neural network(PINN) steady-state prediction model based on the constraints of a composite objective function. The methodology is structured into two primary stages to ensure both data fidelity and physical rigor. First

considering the highly non-linear and coupled operational characteristics of the desulfurization process

the maximal information coefficient(MIC) method is employed to filter six core features from a high-dimensional candidate input space. This approach effectively eliminates dimensional redundancy and ensures that the selected variables(such as slurry pH and circulation flow rates) possess the strongest statistical and physical relevance to the outlet SO

2

concentration. Second

the core innovation lies in the seamless integration of domain expertise into the deep learning architecture. During the training process

leveraging automatic differentiation(AD) technology and the rectified linear unit(ReLU) function

monotonic physical priors are mathematically reformulated as directional penalty terms within the loss function. These terms act as soft constraints that strictly supervise the feasible solution space of the neural network

guiding the model to learn mapping relationships that remain consistent with established gas-liquid mass transfer mechanisms. This mechanism effectively prevents the model from generating non-physical fluctuations even when faced with noisy industrial data. [Results

]

Experimental validation was conducted utilizing real-world operational datasets from a 600 MW thermal power unit. The results demonstrate that the PINN model exhibits superior performance compared to traditional neural network(NN) model. Specifically

the root mean squared error(RMSE) and mean absolute percentage error(MAPE) on the testing set were markedly reduced from 1.105 5 mg/m

3

and 6.98% to 0.645 4 mg/m

3

and 4.98%

respectively. [Conclusion

]

In conclusion

the proposed prediction framework successfully breaks the limitations of conventional black-box models without increasing the topological complexity of the network. It achieves a significant improvement in both empirical prediction accuracy and physical mechanism consistency

providing a reliable model foundation for the fine-tuned operation and intelligent optimization control of desulfurization systems in large-scale power plants.