Prediction of mass loss rate of multi-source hydrocarbon pool fires based on deep learning

Once an accidental fire occurs in a liquid fuel storage tank, it is easy to produce a domino effect, triggering fires in multiple tanks and posing serious challenges to the safety of the liquid storage tank system. To address this issue, this study systematically investigates the influence of crosswind speed and the number of fuel pools on burning rates through a series of scaled experiments. Using experimental data, a dataset of 12,680 annotated flame images and corresponding burning rates was generated. The improved VGG16 algorithm, along with LeNet and AlexNet models, were trained with experimental data under different wind speeds and pool numbers. The result shows a strong intrinsic correlation between the burning rate and characteristics such as flame morphology, burned area size, and brightness. Among the models, VGG16 demonstrated the best performance with an accuracy exceeding 85% under varying wind speeds and fire source numbers. The improved model accurately predicts the transient burning rate in this paper and previous studies. The improved VGG16 algorithm proposed in this paper has stable performance in predicting various testing conditions, can further improve the level of fire safety toughness, and is expected to provide a reference for the safety of chemical park systems.