Dynamic probability analysis on accident chain of atmospheric tank farm based on Bayesian network

Abstract

Due to massive flammable chemical materials and dense layout, domino accident chain always occurs in major accidents of atmospheric storage tank farm. In this paper, 136 domino effect accidents of atmospheric tanks are analyzed to achieve relative probability by ETA method. Based on accident probability, an analysis model is proposed to define the most likely accident chain and the key tank in the atmospheric tank farm. To get the real accident scenario, the method adopts the heat radiation model of pool fire under the fire extinguishing conditions. Combined with the dynamic Bayesian network, the escalation model of accident chain is established to figure out the dynamic probability of the tank under the condition of fire extinguishing. Finally, an example is used to verify the feasibility of the method. It is of great theoretical significance and application value for the treatment of domino accident in atmospheric tank farm.

Introduction

Massive flammable and volatile liquid materials are stored in chemical atmospheric tank, the maximum volume has reached 150,000 m³ in China, which has significant fire and explosion risk (Ning, 2019). Due to the dense layout, once a fire or explosion occurs, it may cause the domino effect accident of tanks. On March 18, 2019, a naphtha tank fire broke out at ITC tank farm in Houston, USA, triggered domino effect involving 11 tanks in the tank farm to be burned down, the whole accident chain lasted 70 h (CSB, 2020). On June 2, 2013, the explosion and fire of toluene tank in the triphenyl tank farm of Sinopec Dalian Petrochemical Company initialized domino effect, and three other tanks in the tank farm exploded and burst into flame, resulting in four deaths and RMB 6.97 million economic loss. From the above accident cases, domino accidents in the chemical and process industry may cause serious economic losses and casualties. Therefore, assessing the accident chain risk has extremely important theoretical significance and application value for the prevention and control of the domino effect.

Due to the low probability of domino accident, it is difficult to adopt the experimental method to simulate the accident scenario. Summary and analysis of characteristics through accident cases is the main research direction at present. Darbra et al. (2010) made statistics on the accident types, causes, consequences of 225 domino accidents, and found that Explosion-Fire, Fire-Explosion and Fire-Fire are the most common accident chain modes. Abdolhamidzadeh et al. (2011) summarized 224 domino accidents of chemical process equipment, explosion is main primary accident, accounting for 57%.

Analyzing the escalation probability and evaluating the severity of the accident with the help of numerical simulation software or the empirical model is an important part of the risk assessment of the domino effect. Khan and Abbasi (1998a) built a model to evaluate the scale and possible consequences of the domino effect in chemical process industries. And they developed the DOMIFFECT software, which can evaluate the domino effect and can be used for the possibility of domino accidents in chemical process industries (Khan and Abbasi, 1998b). Cozzani and Salzano (2004) modified probit model under overpressure damage and proposed a simplified escalation probability model (Salzano and Cozzani, 2005a) and damage threshold of atmospheric tank (Cozzani et al., 2005b, Cozzani et al., 2006). Mingguang and Juncheng (2008) improved the accuracy of the overpressure failure probability model. Abdolhamidzadeh et al. (2010) proposed a Monte Carlo-based algorithm to evaluate the domino effect of chemical process accidents. Landucci et al., 2009a, Landucci et al., 2009b analyzed the failure time of storage tanks under the influence of fire and improved the probit model under the effect of heat radiation. After that, Landucci et al. (2015) researched the accident case data with safety barrier function and improved the calculation model of device failure time. Zhang et al., 2019a, Zhang et al., 2019b analyzed 165 domino accidents, established the propagation probability model of the accident chain, and verified the accuracy of the model.

In recent years, some scholars have begun to use Bayesian network to assess the risk of the domino effect. Khakzad et al. (2013) applied Bayesian network to the study of domino accident and determine the most likely escalation path of domino effect in tank farm. Khakzad, 2015, Khakzad et al., 2018 introduced dynamic Bayesian (DBN) method to simulate the accident escalation in domino scenario, which made up for the shortage of static Bayesian in time escalation. Zeng et al. (2020) combined with dynamic Bayesian network, explained the changes in the escalation probability of domino accidents under the protection of safety barriers. WenHu established an accident probability model of hybrid storage tank area. The paper analyzed and calculated the damage probability of explosion and pool fire to adjacent equipment in liquefied gas storage tanks by considered that the occurrence of a pool fire accident in a single tank is a result of the incident (WenHu et al., 2019). Zerouali and Hamaidi (2019) combined bow tie and fuzzy logic with Bayesian networks to propose a fire and explosion risk prediction model for storage tanks, and verified the effectiveness of the proposed method. Santana combines fuzzy mathematics with Bayesian networks to propose a new method to determine the failure probability of thermal radiation.

Although many scholars have studied the characteristics of domino accidents in the tank farm, specific research of domino accidents on the types of tanks is lacking, as atmospheric tanks and pressurized tanks are different in terms of containing chemicals, internal pressure, and tank structure. In previous literature, the escalation probability model does not consider the real extinguishing process of fire in the propagation process of accident chain, which may last for hours before the fire burnt out on. Therefore, the dynamic probability analysis method of domino accident chain in atmospheric tank farm is proposed in this paper. Event trees are constructed to achieve prior probability through 136 domino accidents. Then the equivalent pool fire radius is proposed to analysis the influence of fire-fighting factors on the escalation probability. Finally, the method is applied to a case study.

The method proposed to assess the risk of domino effect in tank farms and determine the most likely accident chain and key accident tank, which can provide guidance for the layout design and emergency response of atmospheric tank farm.