A study on improving temperature field prediction accuracy using a surrogate model assisted by airflow field

• Surrogate Model: A global surrogate model is proposed for temperature field prediction in coal-fired boiler furnaces, overcoming the limitations of traditional CFD simulations. • Wind Velocity Data Integration: The model uses wind velocity field data, improving prediction accuracy by aligning it with temperature field dimensions through radial basis function interpolation. Notably, the ResNet model performs predictions in approximately 0.386 seconds. • Synthetic Data Generation: Convolutional variational autoencoders (CVAE) generate synthetic data to expand the dataset, enhancing model generalization. • Time Efficiency: The deep learning model drastically reduces prediction time from hours to milliseconds, with ResNet completing predictions in about 0.386 seconds. • Real-Time Application: The surrogate model supports real-time temperature predictions, suitable for intelligent boiler control. Due to the constraints of current measurement methodologies, there is a paucity of data regarding the temperature distribution within coal-fired boiler furnaces. While finite element analysis (FEA) offers high accuracy, its extensive computational requirements make it impractical for real-world applications. This paper presents a global surrogate model incorporating wind velocity field data to overcome this limitation and enable efficient temperature field prediction. Wind velocity, a key variable, is inherently linked to the dynamic behavior of the temperature field. By utilizing radial basis function interpolation, the wind data is transformed to match the dimensions of the temperature field, thereby improving the model's ability to predict complex temperature distributions. Additionally, convolutional variational autoencoders (CVAE) are employed to generate high-quality data samples under varying operational conditions, enlarging the dataset and enhancing the model's generalization capacity. This study develops five deep-learning surrogate models—ResNet, AlexNet, VGG, SegNet, and U-Net—to predict temperature fields. Experimental results demonstrate that including wind velocity data significantly improves prediction accuracy, with the ResNet model achieving an average error of just 1.275%. Compared to traditional computational fluid dynamics (CFD) simulations, the deep learning surrogate model reduces prediction time to milliseconds. The ResNet model surpasses other models, balancing accuracy and computational efficiency, delivering more precise temperature field predictions with lower computational costs. The proposed surrogate model exhibits strong generalization ability and real-time prediction potential, providing valuable support for intelligent boiler control and optimization.