Analysis of vapor cloud explosion behavior in LPG spherical tanks using multi-energy modeling and machine-learning-based factor assessment.

Abstract

Vapor cloud explosions (VCEs) associated with liquefied petroleum gas (LPG) storage systems represent a significant hazard in refinery operations. Accurate estimation of explosion distance is essential for safety setback design and quantitative risk assessment. This study developed an integrated framework that combines physics-based and data-driven techniques to study VCE behavior resulting from LPG releases from a 20,000-barrel pressurized spherical tank under realistic refinery conditions. A total of 336 leakage scenarios were simulated using DNV PHAST 2022 and the Multi-Energy method. Seven influencing parameters were considered, including leak diameter, LPG composition (propane-butane ratio), leak location, atmospheric category, seasonal period, day/night conditions, and three overpressure thresholds (0.02, 0.14, and 0.21 bar). The results revealed a wide dispersion in explosion distance across scenarios, with maximum distances exceeding 2,100 m for catastrophic full-bore ruptures. Feature-importance analysis using a Random Forest regression model showed that leak diameter was the dominant controlling parameter, accounting for approximately 74% of the predictive importance, while LPG composition contributed an additional 24-25%. Operational and atmospheric parameters had comparatively minor effects. The three investigated overpressure thresholds produced nearly identical explosion distances across the evaluated scenarios, indicating that within the low-pressure range, explosion extent is primarily governed by the released fuel mass and flammable cloud geometry rather than the threshold value. A classical power-law correlation was used to describe the geometric scaling between leak diameter and explosion distance; however, its predictive capability was limited (RMSE ≈ 337 m) due to its single-variable formulation. In contrast, the Random Forest model captured multi-factor interactions within the PHAST-generated dataset with very high predictive accuracy (R² ≈ 0.9997, RMSE ≈ 9 m). This hybrid framework provides a transparent approach for analyzing LPG vapor cloud explosions and offers practical insights for safety-distance determination, refinery layout optimization, and risk-based inspection planning.