Probabilistic real-time natural gas jet fire consequence modeling of offshore platforms by hybrid deep learning approach

Abstract

Natural gas jet fire induced by igniting blowouts has the potential to cause critical structure damage and great casualties of offshore platforms. Real-time natural gas jet fire plume prediction is essential to support the emergency planning to mitigate subsequent damage consequence and ocean pollution. Deep learning based on a large amount of Computational fluid dynamics (CFD) simulations has recently been applied to real-time fire modeling. However, existing approaches based on point-estimation theory are 'over-confident' when prediction deficiency exists, which reduce robustness and accuracy for emergency planning support. This study proposes probabilistic deep learning approach for real-time natural gas jet fire consequence modeling by integrating variational Bayesian inference with deep learning. Numerical model of natural gas jet fire from offshore platform is built and the natural gas jet fire scenarios are simulated to construct the benchmark dataset. Sensitivity analysis of pre-defined parameters such as MC (Monte Carlo) sampling number m and dropout probability p is conducted to determine the trade-off between model's accuracy and efficiency. The results demonstrated our model exhibits competitive accuracy with R² = 0.965 and real-time capacity with an inference time of 12 ms. In addition, the predicted spatial uncertainty corresponding to spatial jet fire flame plume provides more comprehensive and reliable support for the following mitigation decision-makings compared to the state-of-the-art point-estimation based deep learning model. This study provides a robust alternative for constructing a digital twin of fire and explosion associated emergency management on offshore platforms.

Keywords: Deep learning; Digital twin; Offshore platform; Real-time jet fire modeling; Variational Bayesian inference.