Understanding the dynamic burning behavior of pool fires under different ullage heights is critically important for process safety and environmental protection, as ullage height directly governs burning rates, flame morphology, and associated heat transfer mechanisms. We propose a deep learning framework to integrate temporal image characteristics with structural scale parameters to predict burning rates throughout the entire combustion process. We constructed an experimental dataset covering diverse combinations of pool diameters and ullage heights, enabling comprehensive evaluation across multiple fire scenarios. The model combines the convolutional neural network (CNN)’s ability to extract spatial features from flame images with the transformer’s strength in capturing long-range temporal dependencies, while incorporating structural parameters through multimodal fusion. Comparative results against several representative deep learning architectures demonstrate that the CNN-Transformer achieves superior performance in terms of absolute error–based metrics and exhibits enhanced robustness in capturing transient and steady burning-rate dynamics, indicating strong predictive accuracy and generalization capability. These findings provide not only a theoretical basis but also practical technological support for quantitative risk assessment, intelligent monitoring, and early warning of pool fires, offering new insights into data-driven approaches for fire safety management in industrial processes.
Gao et al. (Sun,) studied this question.