Situational Awareness for Oil Storage Tank Accidents Based on Complex Networks and Evidence Theory

Abstract

To address the difficulty frontline commanders face in accurately perceiving fireground risks during the early stages of oil storage tank fires, in this study, we propose a method that integrates complex network theory with a multi-source information fusion approach based on cloud models and Dempster-Shafer (D-S) evidence theory for situational analysis and dynamic perception. Initially, the internal evolution of accident scenarios within individual tanks is modeled as a single-layer network, while scenario propagation between tanks is represented through inter-layer connections, forming a multi-layer complex network for the storage area. The importance of each node is evaluated to assess the risk level of scenario nodes, enabling preliminary situational awareness, with limited reconnaissance information. Subsequently, the cloud model’s capability to handle fuzziness is combined with D-S theory’s strength in fusing multi-source data. Multi-source heterogeneous information is integrated to obtain the confidence levels of key nodes across low, medium, and high-risk categories. Based on these results, high-risk scenarios in oil storage tank emergency response are dynamically adjusted, enabling the updating and prediction of accident evolution. Finally, the proposed method is validated using the 2015 Gulei PX plant explosion case study. The results demonstrate that the approach effectively identifies high-risk scenarios, enhances dynamic situational perception, and is generally consistent with actual accident progression, thereby improving emergency response capability.

1. Introduction

As the petrochemical industry continues to expand, oil storage tank areas are increasingly evolving toward regionalization and integration. By 2023, China had approximately 35,000 petrochemical storage tanks in operation [1]. However, highly concentrated layout results in minimal spacing between tanks and creates intricate interdependencies among them. Given the multitude of influencing factors, even a minor failure at any point in the system can trigger a chain reaction, leading to secondary incidents such as fires or explosions. Interaction among multiple hazardous sources within an industrial park can rapidly expand the scope of an accident, potentially resulting in catastrophic consequences [2,3], as evidenced by 7.16 Dalian oil pipeline explosion accident [4] and oil tank on fire in Dongxing Industrial Zone, Nandagang Industrial Park, Cangzhou, Hebei Province [5]. These incidents pose significant challenges for emergency firefighting and rescue operations. Therefore, accurately understanding the accident dynamics in oil tank areas is critical for frontline incident commanders to make timely and effective decisions during emergency response.

In recent years, situational awareness has been increasingly applied in fields such as process safety. Within emergency response operations, frontline commanders—who serve as key tactical decision-makers—must accurately assess and interpret the scene of a storage tank accident. Their ability to quickly perceive the evolving fireground situation directly determines the effectiveness and efficiency of the entire firefighting and rescue mission [6,7]. Based on this, scholars have investigated situational awareness from multiple perspectives. Sheng et al. [8] introduced a UAV (unmanned aerial vehicle)- based method for real-time fire situational awareness, enabling dynamic prediction of failure times for nearby equipment exposed to fire. Zhou et al. [9] developed an event-driven modeling approach based on state transition mechanisms to support situational awareness and decision-making in fire-induced domino emergency scenarios. Naderpour et al. [10] applied the concept of situational awareness to process system safety and developed an innovative approach for situational risk perception by integrating dynamic Bayesian networks with fuzzy set theory. Dos Santos et al. [11] conducted a goal-directed task analysis, providing insights into situation awareness evaluation, situation awareness-oriented operations, firefighter training, and supportive technologies for all-hazard responses. Luokkala et al. [12] examined differences in situational understanding and emergency response capabilities among various decision-making groups during unexpected events.

Effective situational awareness on the fireground first requires a thorough understanding and analysis of fire accidents. Researchers have applied a wide range of computational and mathematical models—such as Bayesian networks, complex networks, neural networks, Petri nets, bow-tie analysis, and artificial intelligence—to accident analysis [13,14,15,16,17]. Among these, complex network models characterized by intricate topological structures, offer a clear representation of causal relationships in accident evolution. They are particularly useful for capturing dynamic interactions among multiple factors, thereby revealing how individual components influence system behavior and overall functionality. As a result, complex networks have become a powerful tool for accident analysis, enabling researchers to uncover hidden connections and systemic vulnerabilities. Feng et al. [18] developed a dynamic accident propagation model tailored to complex industrial systems, revealing the pathways through which incidents spread within network structures and how risks are transmitted. Li et al. [19] applied complex network theory to establish an analytical and control framework for the evolution of fire risk in oil storage tanks, providing a theoretical foundation for fire prevention and mitigation. Liu et al. [20] utilized association rule mining to uncover intrinsic correlations among various accident factors in oil tank areas and used a complex network model to identify key contributing elements. In another study, Li et al. [21] systematically examined how functional modules within chemical enterprises contribute to major accidents, enhancing overall process safety by modeling these relationships through complex networks. Yang et al. [22] constructed a causal network for chemical accidents and identified critical risk nodes using a causal reasoning approach, proposing targeted risk control strategies accordingly.

During actual emergency operations, the emergence of multi-source heterogeneous data places higher demands on the ability to process uncertain information. Several approaches are commonly employed for uncertainty modeling, including probability theory, Dempster-Shafer (D-S) evidence theory, possibility theory, fuzzy set theory, and rough set theory [23,24]. Among these, D-S evidence theory demonstrates significant advantages in scenarios in which prior knowledge is limited. It allows for the effective integration of multi-source information and helps to reduce overall system uncertainty, making it a promising tool for practical applications in emergency environments [25]. Durukan et al. [26] integrated D-S evidence theory with Hazard and Operability analysis and fault tree methodology to develop a risk assessment framework for predicting fire and explosion hazards caused by failures in inert gas systems aboard tankers. By integrating D-S evidence theory with game theory, Xia et al. [27] proposed a sequential decision-making framework aimed at anticipating situational changes at key locations during major accidents, thereby strengthening the scientific rigor and predictive capability of emergency response efforts. Niu et al. [28] established a fire risk evaluation model for urban utility tunnels by incorporating an improved cloud model with D-S evidence theory. Wang et al. [29] designed a multi-sensor data fusion approach based on modified D-S theory and demonstrated its effectiveness in detecting tunnel fires.

In summary, current research on situational awareness in disaster accidents primarily focuses on pre-incident risk assessment and early warning, while support for post-incident emergency command and decision-making remains limited. To address this gap, in this study, we aim to help realize situational awareness during fire incidents based on the practical needs of frontline fire commanders. Complex networks can reveal the systemic relationships within accident evolution, while D-S evidence theory offers distinctive advantages in handling the uncertainty of multi-source fireground information. We employ a multi-layer complex network approach to dynamically characterize the fire evolution process. In the early stages of a fire, the model constructs accident chains based on historical data and identifies critical scenarios by analyzing key nodes within the complex network, thereby achieving preliminary situational awareness. As real-time information from the scene becomes available, the integration of the cloud model and D-S evidence theory enables the fusion of multi-source data, significantly enhancing real-time understanding and prediction of fireground conditions. Ultimately, this approach provides more timely and targeted decision support for frontline emergency responders.

2. Methods

2.1. Complex Network Node Assessment Metrics

Complex networks can reveal the evolutionary dynamics of interactions among multiple factors, thereby explaining the overall functioning of complex systems and the influence of individual components on these systems. Research on network structures can identify critical nodes [30], thereby enhancing the understanding of such interactions.

Identifying key nodes in a network depends on assessing the importance of each node. Centrality is a core indicator used for this purpose, typically evaluated using degree centrality, betweenness centrality, and closeness centrality [31]. These metrics capture the influence of each tank within the overall progression of a fire scenario.

Degree centrality reflects the number of direct connections a node has within the network. In an oil tank accident scenario, nodes with high degree centrality typically exert greater influence over the structure and propagation of the event. The calculation is shown in Equation (1):

$$D_c(i)=\sum _{j\ne i\in N}{\displaystyle \frac{{\delta }_{ij}}{n-1}}$$

(1)

where $D_c\left(i\right)$ denotes the degree centrality, ${\delta }_{ij}$ represents the cardinality of directly adjacent nodes, and $n$ indicates the total node count in the network.

Betweenness centrality measures how often a node functions as a bridge along the shortest path between two other nodes. In a fire propagation scenario, nodes with high betweenness centrality exert greater control over the spread of the event. The calculation is shown in Equation (2):

$$D_b(i)=\sum _{s\ne i\in N}{\displaystyle \frac{{\sigma }_{st}\left(i\right)}{{\sigma }_{st}}}$$

(2)

where $D_b\left(i\right)$ denotes betweenness centrality. ${\sigma }_{st}\left(i\right)$ expresses the number of shortest paths between nodes s and t that pass through the focal node i, while ${\sigma }_{st}$ stands for the total number of shortest paths between nodes s and t.

Closeness centrality indicates how quickly a node can influence other nodes following an incident. Higher closeness centrality means that the node more efficiently spreads its influence throughout the network. This is described in Equation (3):

$$D_{cc}(i)={\displaystyle \frac{n-1}{\sum _{j\ne i\in N}{d}_{ij}}}$$

(3)

where $D_{cc}\left(i\right)$ denotes closeness centrality, $d_{ij}$ represents the shortest path distance between nodes i and j, and $n$ indicates the total number of nodes in the network.

Evaluating node importance using only one centrality measure does not adequately identify critical nodes. Therefore, we calculate all three metrics—degree, betweenness, and closeness centrality—and perform a comparative analysis. A weighted decision matrix is constructed to assess the overall importance of each node in the network, as expressed in Equation (4):

$$Z_i=\left(\sum _j{E}_j{\displaystyle \frac{K_{ij}}{G_{ij}}}\right)\times 100\%$$

(4)

where $Z_i$ denotes the importance of node i, and $E_i$ represents the weight of metric j. In this study, three distinct node metrics are selected, and each metric is thus assigned an equal weight of 1/3. $K_{ij}$ is the value of metric j for node j, and $G_{ij}$ is the average value of metric j across all nodes in the complex network.

The quantified importance of a node reflects its role in the network layer and helps estimate its risk level in the context of a fire accident. Generally, a node with higher importance contributes more significantly to accident propagation and is assigned a higher risk level.

2.2. Cloud Model Theory

Cloud models are mathematically established frameworks based on statistical mathematics and fuzzy mathematics, characterizing the randomness and fuzziness between uncertain linguistic values and precise numerical values [32]. Let $U=\{u_1,u_2,\dots ,u_n\}$ be a quantitative set represented by precise numerical values, referred to as the domain of discourse. Let $Q$ denote a qualitative concept defined over $U$. For any variable, $x$, there exists a random number, $\mu \left(x\right)\in \left[\mathrm{0,1}\right]$, with a stable tendency, representing the degree of certainty of $x$ with respect to the concept $Q$. The distribution of $x$ over the universe $U$ is referred to as a cloud, and the process of generating cloud drops describes the uncertainty mapping between qualitative concepts and quantitative values.

$$\mu (x)=e^{{\displaystyle \frac{(x-Ex{)}^2}{2{\left({\mathrm{E}\mathrm{n}}^{\prime }\right)}^2}}}$$

(5)

where $x\sim N(Ex,{En}^{\prime 2})$ and ${En}^{\prime }\sim N(En,H{e}^2)$.

This model features three numerical characteristics: expectation ($Ex$), entropy ($En$), and hyperentropy ($He$). The cloud model characterizes the holistic features of qualitative concepts through these three numerical parameters, thereby enabling the conversion between qualitative concepts and quantitative expressions [33,34]. For each risk level associated with scenario-specific parameters, a threshold interval is defined with lower and upper bounds denoted as $\left[C_{ij}\right(min), C_{ij}(max\left)\right]$.

(1)

The calculation method for $Ex$ is shown in Equation (6):

$$Ex={\displaystyle \frac{C_{ij}\left(max\right)+C_{ij}\left(\mathrm{m}\mathrm{i}\mathrm{n}\right)}{2}}$$

(6)

(2)

The calculation method for $En$ is shown in Equation (7):

$$En={\displaystyle \frac{C_{ij}\left(max\right)-C_{ij}\left(\mathrm{m}\mathrm{i}\mathrm{n}\right)}{6}}$$

(7)

(3)

The calculation method for $He$ is shown in Equation (8):

$$He= C$$

(8)

where $C$ is an empirically determined value based on expert judgment. To keep the uncertainty within an acceptable range during the identification of risk threshold intervals, the hyper-entropy value is set to 0.05.

2.3. Dempster-Shafer Evidence Theory

D-S evidence theory is one of the most critical techniques in the field of information fusion. It offers greater flexibility in representing uncertainty and accurately reflecting the process of evidence collection.

In D-S evidence theory, the frame of discernment $\Theta$ represents a set of mutually exclusive and collectively exhaustive elements, as shown in Equation (9):

$$\Theta =\{{\theta }_1,{\theta }_2,\dots ,{\theta }_n\}$$

(9)

where ${\Theta }_i$ denotes an event within $\Theta$, and $n$ represents the number of elements in the frame.

The basic probability assignment (BPA) is defined as a mapping from the power set $2^{\mathsf{\Theta }}$ of the frame of discernment $\Theta$ to the interval $\left[0,1\right]$, expressed as $\mathrm{m}:2^{\mathsf{\Theta }}\to \left[\mathrm{0,1}\right]$. Here, $m\left(A\right)$ denotes the basic probability assigned to the subset $\mathrm{A}$ of $\mathsf{\Theta }$. This mapping must satisfy Equations (10) and (11):

$$m(\mathsf{\varnothing })=0$$

(10)

$$\sum _{A\subseteq \Theta }m(A_j)=1$$

(11)

where $m$ represents the mass function, and $A$ is a proposition within $\Theta$, referred to as a focal element. Equation (10) indicates that no belief is assigned to the empty set, and Equation (11) ensures that the sum of all basic probability assignments over the power set equals 1.

If there are two independent basic belief assignments, ${\mathrm{m}}_1$ and ${\mathrm{m}}_2$, in $\mathsf{\Theta }$, then the combination rule of evidence theory is expressed as $m=m_1\oplus m_2$, as shown below.

$$m(A)=\left\{\begin{array}{cc}{\displaystyle \frac{1}{1-K_R}}{\displaystyle \sum _{B\cap C=A}}{m}_1(B)m_2(A),& A\ne \mathsf{\varnothing }\\ 0,& A=\mathsf{\varnothing }\end{array}\right.$$

(12)

$$K=\sum _{B\cap C=\varnothing }{m}_1(B)m_2(A)$$

(13)

in this equation, $B,C\in 2^{\Theta }$ and $K$ represents the conflict coefficient between $m_1$ and $m_2$, which indicates the degree of conflict between the two pieces of evidence. A $K$ value closer to 1 implies a higher level of inconsistency between the information sources. The Dempster-Shafer combination rule is applicable only when the condition $K<1$ is satisfied.

3. Results

3.1. Constructing a Multi-Layer Network Model of Oil Storage Tank Accident Chains

In this section, we construct a complex network based on historical accident cases and analyze the risks of accident scenarios using node indicators within the network. The proposed method is designed for situations in which emergency responders have just arrived at the scene and have not yet collected sufficient information but already possess a preliminary understanding of the overall fire scenario. By establishing a multi-layer network model, the approach enables the identification of critical nodes in the accident, assisting responders in quickly locating and prioritizing high-risk events for timely intervention.

3.1.1. Identification of Oil Storage Tank Accident Scenarios

Fire scenarios involving oil storage tanks play a key role in emergency decision-making. They also serve as a foundation for developing effective response strategies. We analyzed 101 representative accident cases. These were collected from official government sources, such as investigation reports and bulletins issued by the Ministry of Emergency Management of the People’s Republic of China, as well as from the peer-reviewed literature, professional publications, and authoritative media. Based on this analysis, accident scenarios were identified, as detailed in

Table 1. Accident scenarios in oil storage tank areas.

Node ID	Scenario Node	Node ID	Scenario Node
S1	Tank explosion	S7	Sealing ring fire
S2	Boil over	S8	Full surface fire
S3	Tank collapse	S9	Enclosed combustion
S4	Tank rupture	S10	Torch fire
S5	Floating roof failure	S11	Re-ignition
S6	Flowing fire

.

The scenario nodes listed in Table 1 do not necessarily refer to a single, specific type of accident. For example, node S₁ (tank explosion) encompasses explosions caused by a variety of mechanisms, including physical and chemical reactions. Node S₅ (floating roof failure) refers to structural malfunctions such as tilting, jamming, or sinking of a floating roof under fire-related conditions. Node S₉ (enclosed combustion) refers to fire scenarios in confined areas where specific on-site conditions prevent direct application of water or foam by first responders, creating extinguishing blind spots.

3.1.2. Multi-Layer Network Construction for Oil Storage Tanks

In analyzing complex networks of accidents in oil storage tank areas, each major accident can be broken down into multiple distinct Scenarios. These scenarios are treated as nodes in a network, while the relationships between them serve as edges. Accordingly, the oil tank fire accident network can be defined as a binary tuple, V = (G, E), where V represents the set of scenario nodes and E denotes the set of edges capturing relationships between them.

Once the complex network for a single tank incident is constructed, networks for multiple incidents involving the same tank type can be overlaid. This aggregation produces an evolution map representing the typical propagation patterns for that specific tank type. The specific process is shown in Figure 1.

Considering the common tank types-—fixed roof, inner floating roof, and outer floating roof tanks—we construct separate evolution networks for each, based on case analysis. These networks are illustrated in Figure 2.

In the multi-layer complex network model proposed in this study, each layer represents the accident evolution process of an individual storage tank. The nodes correspond to potential accident scenarios within each tank, and the number of nodes in each layer may vary depending on the specific case. During modeling, the multi-layer network is constructed by layering single-tank accident networks (as shown in Figure 3), based on the progression stages and scenario dependencies of different tanks. It should be emphasized that Figure 3 is provided for illustrative purposes only.

The number of nodes in each layer is dynamically adjusted according to the specific accident scenarios, tank type, and progression of the incident. Intra-layer edges describe the evolution of accident scenarios within a single tank, while inter-layer edges represent risk transmission pathways between scenarios in different tanks. Inter-layer connections are determined based on temporal relationships and physical propagation paths between tanks, as reported in accident investigations and documented case studies. For example, after the rupture of Tank 1 causes flowing fire, the resulting thermal radiation affects the adjacent Tanks 2 and 3, triggering secondary incidents. Based on this causal propagation—from one tank accident scenario (such as flowing fire) to another (such as sealing ring fire)—corresponding inter-layer links are established in the multi-layer network.

In real fireground conditions, the initially ignited tank not only undergoes internal evolution but also continuously influences nearby burning tanks, thereby increasing the complexity of the overall incident. Given the limited availability of firefighting resources, emergency responders tend to prioritize burning tanks and their immediate surroundings in order to contain fire spread and optimize resource allocation. Based on these considerations, this study focuses on interactions between burning tanks and adjacent ones. The storage area is structured into three layers: Layer 1 includes initially ignited tanks; Layer 2 includes tanks ignited due to exposure to Layer 1; and Layer 3 comprises tanks affected by fire but not yet ignited.

Within this framework, tanks at the same level may influence one another, but higher-tier tanks do not directly impact lower-tier tanks. Only the evolution of accidents in lower layers may affect those above. This is due to the earlier ignition time of lower-layer tanks, which accelerates their risk propagation compared to that of higher layers. In real fire scenarios, the proposed multi-layer network model can be further extended. For instance, a burning tank in Layer 2 may be further subdivided based on burn duration or environmental conditions. In this case, a long-burning tank may exert influence on a newly ignited one, while an upwind tank may affect those located downwind. This refinement allows the model to capture more complex patterns of risk propagation. Moreover, as the fire spreads, the interactions among tanks evolve over time. Therefore, the network topology should be dynamically updated to reflect real-time fireground conditions, ensuring consistency between risk analysis and actual developments.

3.2. Information Interpretation and Situational Awareness for Major Accidents in Oil Storage Tank Areas

In the previous section, the comprehensive importance of accident scenario nodes in oil storage tank areas was evaluated using multi-layer network indicators, leading to the identification of scenarios requiring attention during emergency response. Given the structural, operational, and environmental variations across different storage tank areas, enhancing the ability to predict the evolution of major accidents requires timely collection of multi-source parameter information closely related to accident progression after an incident occurs, followed by data fusion to support situational assessment.

3.2.1. Identification of Key Scenarios and Associated Information Parameters

Key accident scenarios refer to events that play a decisive role in the evolution direction and final outcome of an accident [35]. While a broad set of scenarios can be initially extracted, only a subset play a decisive role in the evolution of major accidents. Identifying these critical scenarios enables emergency responders to prioritize the collection of relevant parameter information, thereby improving the efficiency and accuracy of situational assessment.

To identify the critical scenarios that may lead to catastrophic consequences in oil storage tank fires, we systematically reviewed and categorized typical accidents based on 101 publicly available major accident cases to ensure comprehensiveness. During the identification process, several factors were considered, including potential casualties caused by delayed response and the difficulty of emergency handling. In addition, three selection criteria were established to enhance objectivity:

• Loss of control over the scenario state leads to significant casualties;
• Loss of control over the scenario state contributes to the escalation of the disaster;
• Loss of control over the scenario state significantly increases difficulties in handling the disaster.

Furthermore, the selection was supported through expert consultation and cross-validation with those with frontline firefighting experience. For example, torch fire is generally considered a relatively stable form of combustion and poses lower risk. In contrast, flowing fire often results from structural failure of storage tanks, due to which leaked oil rapidly ignites upon contact with a fire source. Therefore, the preceding scenarios—such as tank collapse or rupture—deserve greater attention. Moreover, although flowing fire is highly challenging to manage during emergency response, it typically ignites immediately upon encountering a fire source following structural failure. As such, it is difficult to characterize using specific informational parameters. Therefore, flowing fire was not included as a key accident scenario in this study. Based on these considerations, the preliminary set of accident scenarios was refined and optimized. The identification process is illustrated in Figure 4.

Ultimately, five key scenarios were selected due to their significant impact on the evolution of fire incidents; these are showcased in

Table 2. Key accident scenarios in oil storage areas.

Code	Key Accident Scenario
Y1	Tank explosion
Y2	Boil over
Y3	Full surface fire
Y4	Tank collapse
Y5	Re-ignition

. Given the urgency of on-site conditions, this selection enables frontline responders to focus on critical aspects and prioritize the acquisition of key parameters related to these scenarios.

Based on a systematic statistical analysis of historical data from major oil storage tank accidents, in this study, we further incorporated the practical capabilities of current information detection technologies and the reliability of data acquisition to identify key parameters. National standards, industry regulations, and the existing academic literature were consulted, alongside valuable input from frontline emergency responders. From a multi-dimensional perspective, potential informational elements influencing the judgment of accident scenarios were evaluated. As a result, nine critical parameters with strong representativeness and practical relevance in fire emergency response were identified. Furthermore, the relationships between these parameters and the previously identified critical accident scenarios were established, as detailed in

Table 3. Key parameters for oil storage areas.

Code	Key Parameter
X1	Combustible gas concentration
X2	Pressure
X3	Temperature
X4	Foam coverage ratio
X5	Liquid-level-to-tank-height ratio
X6	Water content
X7	Wind speed in downwind direction
X8	Time exposed to heat radiation
X9	Heat radiation

and

Table 4. Correlation table of key accident scenarios and key parameters for oil storage areas.

	Combustible Gas Concentration	Pressure	Temperature	Foam Coverage Ratio	Liquid-Level -to-Tank -Height Ratio	Water Content	Wind Speed in Downwind Direction	Time Exposed to Heat Radiation	Heat Radiation
Tank explosion	√	√	√
Boil over			√		√	√
Full surface fire							√	√	√
Tank collapse			√		√
Re-ignition			√	√

.

3.2.2. Determination of Threshold Intervals for Information Parameters

Based on the opinions and analytical experience of relevant experts, referring to the Guidelines for Data Collection of Online Monitoring and Accident Early Warning Systems for Major Hazard Sources of Hazardous Chemicals (Tank Areas and Storage Yards) (Trial) issued by the General Office of the State Administration of Work Safety of China, and incorporating insights from field interviews with frontline firefighting personnel, the risk thresholds for key accident scenarios were ultimately determined, as shown in

Table 5. Information parameter thresholds corresponding to critical accident scenarios.

Key Accident Scenario		Grade	Low	Medium	High
	Information Parameters
Tank explosion	Gauge pressure at tank top (°C)		[0.20, 2.10)	[2.10, 5.52)	[5.52, 6.90]
	Tank top temperature (°C)		[40, 60)	[60, 80)	[80, 100]
	Gasoline vapor concentration (%)		[0, 1.9)	[1.9, 3.8)	[3.8, 7.6]
	Diesel vapor concentration (%)		[0, 1.88)	[1.88, 3.75)	[3.75, 7.50]
Boil over	Liquid-level-to-tank-height ratio		[0, 0.5)	[0.5, 0.8)	[0.8, 1]
	Water layer temperature (°C)		[20, 64)	[64, 80)	[80, 100]
	Water content (%)		[0, 0.66)	[0.66, 1.32)	[1.32, 2]
Full surface fire	Downwind wind speed (m/s)		[0, 4)	[4, 6)	[6, 8]
	Duration of thermal radiation exposure (min)		[0, 10)	[10, 20)	[20, 30]
	Intensity of received thermal radiation (kW/m2)		[0, 5)	[5, 10)	[10, 15]
Tank collapse	Tank wall temperature (°C)		[0, 300)	[300, 450)	[450, 600]
	Liquid-level-to-tank-height ratio		[0.8, 1)	[0.5, 0.8)	[0, 0.5]
Re-ignition	Foam coverage ratio (%)		[90, 100)	[70, 90)	[0, 70]
	Oil surface temperature (°C)		[20, 40)	[40, 60)	[60, 80]

.

3.2.3. Data Fusion and Situational Awareness Process

After emergency teams arrive at the site of a major oil storage tank accident, various technical tools are immediately deployed to collect on-site data. These tools include thermal imagers, hazardous chemical detectors, and life detection equipment, which are used to measure temperature levels, detect combustible gas concentrations, and search for trapped individuals in the affected area. Simultaneously, fixed monitoring devices installed around the tank zone are inspected for operational status, and the data recorded by these sensors are retrieved for later use in emergency response.

This process generates a substantial volume of heterogeneous information originating from multiple sources and measured in diverse units—referred to as multi-source on-site data. To process this data effectively, a method combining the cloud model and D-S evidence theory is employed for information fusion and situational understanding.

(1)

First, the key accident scenarios that require focused attention are identified and the risk parameters that are highly relevant to these scenarios are determined, as detailed in Section 3.2.1.

(2)

Second, based on existing accident research findings, national industry standards, and practical feedback from frontline emergency responders, risk threshold intervals are established for the key parameters. These intervals are used to effectively match observed parameter values with predefined hazard levels, thereby enabling an initial classification of risk levels (Section 3.2.2).

(3)

After defining the threshold intervals for each key parameter across different risk levels, the cloud model is applied to process the multi-source data obtained from sensors. Specifically, Equations (6)–(8) are used to calculate the numerical characteristics of each parameter—namely, $Ex$, $En$, and $He$—to represent their statistical behavior across different risk levels. Based on these computed characteristics, the corresponding parameter values for each critical accident scenario are evaluated to determine their membership degrees to different risk levels. These membership degrees are then used as the BPAs under the established recognition framework. The relevant computational formulas are presented as follows.

$$m\left(A\right)=\mu \left(A\right)=\exp \left[-{\displaystyle \frac{{\left(x-\mathrm{E}{x}_{ij}\right)}^2}{2{\left({{En}_{ij}}^{\prime }\right)}^2}}\right]$$

(14)

$$m\left(\Theta \right)=\mu \left(\Theta \right)=1-\sum _{i=1}^N{\mu }_i\left(A_j\right)$$

(15)

$${{En}^{\prime }}_{ij}=E{n}_{ij}+H{e}_{ij}\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{d}( )$$

(16)

$$\mu =\sum _{k=1}^N{\displaystyle \frac{\mu \left(A\right)}{N}}$$

(17)

where $m\left(A\right)$ represents the BPA value of the parameter information at a specific risk level, $m\left(\Theta \right)$ represents the BPA value under uncertainty, $x$ refers to the numerical value of the parameter, $rand( )$ is a random value between 0 and 1, and $N$ denotes the number of cloud drops.

(4)

Finally, the BPAs are dynamically fused using D-S evidence theory to produce a real-time, integrated interpretation of the incident scene. This approach determines quantifiable confidence levels for each key scenario under different risk grades, thus supporting a comprehensive understanding of evolving situations in large-scale oil storage tank accidents.

4. Case Study

On 6 April 2015, at 18:56, an explosion occurred at the Tenglong Aromatics PX project in Gulei, Zhangzhou, Fujian Province, China. This study takes this incident as a case to analyze its initial conditions and analyze the evolution process of fire-related risks. According to field investigations, the explosion originated in the adsorption–separation unit of the facility. The blast subsequently affected four inner floating roof tanks—607, 608, 609, and 610—located in the central tank area to the west of the unit, and their layout is shown in Figure 5.

4.1. Situation Awareness of Oil Storage Tanks Based on Complex Networks

Based on the incident scenarios involving these inner floating roof tanks, a single-layer complex network model was constructed to illustrate the propagation of fire risk, as shown in Figure 2b. In this network, each node represented a specific incident scenario related to an inner floating roof tank, and the edges represented potential causal or risk transmission relationships among these scenarios. This structured captures how risk may propagate within an affected tank system.

In the initial stage, the specific damage status of the four tanks remained unclear. Therefore, they were temporarily treated as being on the same hierarchical level, and their mutual influence was considered, as depicted in Figure 6. As the accident in the tank area progressed, the multi-layer network model was updated accordingly to reflect the evolving risk dynamics. For analytical clarity, the fire scenarios associated with tanks 607, 608, 609, and 610 were labeled as SA, SB, SC, and SD, respectively.

Upon arrival at the scene, frontline emergency responders observed that tanks 607, 608, and 610, all inner floating roof tanks with similar structural characteristics, had almost simultaneously been damaged by the blast and were intensely combusting. Since these tanks experienced the same type of accident and exerted significant impact on tank 609, they were classified as lower-layer nodes in the multi-layer network model, representing units where the fire had already occurred. In contrast, tank 609, which remained structurally intact without visible signs of damage but was at considerable risk due to exposure to the surrounding burning tanks, was classified as a higher-layer node. In addition, based on the on-site meteorological conditions of easterly and northerly winds, the model further incorporated the directional influence of wind on fire spread. Within the lower-level nodes, tank 607 could be regarded as a potential source point according to the wind direction, and it was considered to have a strong influence on the adjacent tanks 608 and 610.

Given the severe structural damage and the intense burning conditions—approaching the characteristics of full surface fires—scenarios such as tank explosion, tank rupture, and sealing ring fires were excluded from the initial single-layer network model for tanks 607, 608, and 610. Based on field reconnaissance and real-time fireground observations, a multi-layer network was constructed to simulate risk propagation across the tank area, as illustrated in Figure 7.

Taking node SA₅ as an example, its degree centrality was calculated as follows: SA₅ was directly connected to 2 other nodes in the constructed network, which contained a total of 23 nodes. Thus, the degree centrality of SA₅ was 0.09.

As for betweenness centrality, SA₅ did not lie on any of the shortest paths between other node pairs; therefore, its betweenness centrality was 0.

Additionally, the total length of the shortest paths from SA₅ to all other nodes was 88. Based on this, the closeness centrality of SA₅ was computed as 0.25, and the overall node importance score was Z = 0.36.

Following the same procedure, the centrality metrics and importance scores of the remaining nodes in the network were calculated by substituting their respective values into the corresponding formulas. The results are shown in Figure 8.

As shown in Figure 8a, node SA₈—representing full surface fire in tank 607—exhibited the highest values across all three measures for this tank: degree centrality, betweenness centrality, and closeness centrality. In comparison, the full surface fire nodes of tanks 608 and 610 (SB₈ and SD₈) also held the highest degree centrality for these tanks, but the scenarios with the highest betweenness and closeness centrality were the flowing fire nodes (SB₆ and SD₆), respectively. For tank 609, the explosion scenario (SC₁) ranked highest in all three centrality metrics, indicating its central role within the network. These findings suggest that the full surface fire and flowing fire scenarios acted as critical nodes for risk propagation within the network. The high centrality of the tank explosion scenario in tank 609 indicates that it was the most vulnerable to cascading risk, potentially triggered by neighboring tanks. The explosion in tank 609 was likely caused by intense thermal radiation from the adjacent tanks undergoing flowing fire and full surface fire. This radiation accelerated the evaporation of internal oil and gas, increasing internal pressure. Once the pressure exceeded the structural tolerance of the tank, an explosion was triggered.

Based on the integrated node importance within the multi-layer network (with the most critical nodes for each tank highlighted in red in Figure 8b), the critical scenario for tank 607 was identified as full surface fire. In this condition, tank 607 not only faced a severe combustion risk itself but also emitted significant thermal radiation, thereby accelerating the propagation of risk to adjacent tanks. For tank 608, the primary risk scenario was flowing fire. Field observations indicated that prolonged thermal radiation compromised the structural integrity of the tank, eventually causing it to collapse. The resulting leak of flammable liquid led to the formation of a flowing fire, which ignited nearby tanks and escalated the incident. Tank 609 remained relatively safe during the initial phase of the incident. However, as surrounding tanks experienced full surface fires, the impact of thermal radiation became significant. This led to intensified vaporization of internal oil and gas, causing a steady rise in internal pressure. Once this pressure exceeded the structural threshold, an explosion became highly likely. The calculated results align with the actual progression of the incident, confirming that tank 609 eventually exploded due to prolonged exposure to high thermal loads. Additionally, the most critical scenario for tank 610 was identified as flowing fire. This is primarily due to its liquid level being just below mid-level and sustained thermal radiation reducing the tank’s structural stability. If rupture was to occur, leaked flammable liquid could quickly form a flowing fire, significantly expanding the fire’s reach and amplifying the severity of the incident.

Therefore, when firefighters first arrived at the scene or when on-site parameter data were difficult to obtain in a timely manner, frontline responders should have given priority to preventing several high-risk scenarios during emergency handling, including the full surface fire of tank 607, the spread of flowing fire from tank 608, the explosion risk associated with tank 609, and potential flowing fire involving tank 610. Corresponding emergency response plans should be developed in advance to minimize further escalation. Specifically, full surface fires may be suppressed via upwind foam application with sufficient discharge intensity, complemented by advanced techniques such as three-phase jet extinguishing. Flowing fires can be mitigated through containment measures (e.g., dikes or embankments) combined with directional foam application to achieve encirclement control. For tanks at risk of explosion, continuous external cooling and depressurization should be maintained to reduce the probability of catastrophic failure.

4.2. Refined Situational Perception of Oil Storage Tanks Using Reconnaissance Information

After completing preliminary risk analysis based on complex networks, key scenarios and high-risk nodes in an accident can be identified. However, due to the often incomplete or delayed availability of on-site information when rescue teams first arrive, relying solely on historical evolution paths introduces a degree of uncertainty. Therefore, it is essential for responders to promptly collect field data, including critical parameters such as temperature and tank pressure. By integrating this multi-source information, the initially identified key scenario nodes can be further validated and refined, enabling a more accurate and dynamic understanding of the current fireground situation.

Based on the correspondence between multi-source information and critical accident scenarios, in this study, we analyzed the field data collected from the accident site. Specifically, the numerical characteristics of each information parameter of the cloud model at various risk levels were calculated using Equations (6)–(8). The detailed results are presented in Figure 9 and

Table 6. Numerical characteristics of information parameters of the cloud model.

Key Accident Scenarios		Grade	Low			Medium			High
	Information Parameters
			$\mathit{E}\mathit{x}$	$\mathit{E}\mathit{n}$	$\mathit{H}\mathit{e}$	$\mathit{E}\mathit{x}$	$\mathit{E}\mathit{n}$	$\mathit{H}\mathit{e}$	$\mathit{E}\mathit{x}$	$\mathit{E}\mathit{n}$	$\mathit{H}\mathit{e}$
Tank explosion	Gauge pressure at tank top (°C)		1.15	0.32	0.05	3.81	0.57	0.05	6.21	0.23	0.05
	Tank top temperature (°C)		50	3.33	0.05	70	3.33	0.05	90	3.33	0.05
	Gasoline vapor concentration (%)		0.95	0.32	0.05	2.85	0.32	0.05	5.70	0.63	0.05
	Diesel vapor concentration (%)		0.94	0.31	0.05	2.82	0.31	0.05	5.63	0.63	0.05
Boil over	Liquid-level-to-tank-height ratio		0.25	0.08	0.05	0.65	0.05	0.05	0.90	0.03	0.05
	Water layer temperature (°C)		42	7.33	0.05	72	2.67	0.05	90	3.33	0.05
	Water content (%)		0.33	0.11	0.05	0.99	0.11	0.05	1.66	0.11	0.05
Full surface fire	Downwind wind speed (m/s)		2	0.67	0.05	5	0.33	0.05	7	0.33	0.05
	Duration of thermal radiation exposure (min)		5	1.67	0.05	15	1.67	0.05	25	1.67	0.05
	Intensity of received thermal radiation (kW/m2)		2.5	0.83	0.05	7.5	0.83	0.05	12.5	0.83	0.05
Tank collapse	Tank wall temperature (°C)		150	50	0.05	375	25	0.05	525	25	0.05
	Liquid-level-to-tank-height ratio		0.90	0.03	0.05	0.65	0.05	0.05	0.25	0.08	0.05
Re-ignition	Foam coverage ratio (%)		95	1.67	0.05	80	3.33	0.05	35	11.67	0.05
	Oil surface temperature (°C)		30	3.33	0.05	50	3.33	0.05	70	3.33	0.05

.

According to the on-site investigation report, tanks 607 to 610 were four inner floating roof tanks, each with a capacity of 10,000 m³. On the day of the fire, the stored volumes were 6622 m³ in tank 607, 1837 m³ in tank 608, 1563 m³ in tank 609, and 4020 m³ in tank 610. Based on these figures, the liquid level–height ratio for each tank could be calculated, which was then used to determine their membership degrees across different risk levels. As tank 609 did not catch fire during the incident, tank 608 was selected as a representative case for data fusion analysis. Given that further real-time field data have not yet been disclosed in the incident report, we referred to the Code for fire protection design of petroleumand natural gas engineering (GB50183-2004) [36], in which the “Commentary” section includes relevant statements indicating that oil tanks—especially above-ground steel tanks—experience rapid temperature increases under fire conditions, with the temperature of the tank wall above the oil surface reaching approximately 500 °C within 5 min of exposure to flames. Therefore, in this study, we used a tank temperature of 500 °C and a liquid–level height ratio of 0.18 as example input values to demonstrate the multi-source information fusion process, aiming to better support fireground situational assessment and response decision-making.

By substituting the above data into Equations (14) and (15), the basic probability assignments (BPAs) for each risk scenario were calculated. This step enabled the transformation of multi-source information into quantifiable support for decision-making. The resulting basic probability distributions for the corresponding scenarios are summarized in

Table 7. Distribution and fusion results of BPAs for the risk of tank collapse.

		Low	Medium	High	Uncertain
Tank collapse	Liquid-level-to-tank-height ratio	0.0000	0.0007	0.5670	0.4323
	Tank wall temperature (°C)	0.0000	0.0000	0.6065	0.3935
	Fusion	0.0000	0.0003	0.8295	0.1702

.

According to the multi-source information fusion results, the membership degree of the high-risk level for tank collapse reached 0.8295, significantly higher than other risk levels, indicating a high probability of tank failure under current conditions. Specifically, the tank wall temperature reached 500 °C, resulting in a high-risk membership degree of 0.5670, with a certain level of uncertainty (0.4323). The liquid-level-to-tank-height ratio was 0.18, also showing a relatively high membership to the high-risk zone (0.6065). After fusing these two key parameters, the identification of high-risk scenarios became more evident. These results suggest that as the fire intensified, the structural deformation risk of the tank wall increased, especially under low liquid levels. Priority should have been given to cooling and structural stabilization measures, with emphasis on maintaining uniform and continuous cooling of the tank wall above the liquid level from a safe distance to preserve structural integrity.

Case analysis shows that complex network theory can be used in the early stages of an accident to construct propagation path models based on historical data. By calculating node indicators and identifying high-risk nodes, key scenarios can be recognized to achieve initial situational awareness. As the incident progresses and more information is collected, cloud models and D-S evidence theory can be introduced to integrate fireground reconnaissance data, improving the accuracy of situational understanding and prediction.

5. Conclusions

This study focuses on situational awareness and emergency decision-making in oil storage tank fire incidents and establishes a situational awareness framework that integrates complex network analysis with multi-source information fusion. The main contributions of this research are reflected in the following three aspects:

(1)

Single-layer network models for fixed roof, inner floating roof, and outer floating roof tanks were first established based on complex network theory. Building on these models, a multi-layer fire propagation framework was developed by incorporating the logical relationships among different accident scenarios across tanks. By integrating degree, betweenness, and closeness centrality, the analysis identified high-risk nodes within accident chains. The results provide preliminary references for firefighters in the early stage of fire response, particularly when real-time fire data are not yet available.

(2)

Through a comprehensive review of the literature, historical accident cases, and frontline firefighter experience, five critical accident scenarios in oil storage tank areas—namely tank explosion, boil over, tank collapse, re-ignition, and full surface fire—were identified and their associated parameters were analyzed. On this basis, a multi-source information fusion mechanism was constructed by integrating cloud models and D-S evidence theory, which enables the fusion of reconnaissance data collected by firefighters upon arrival at a scene. This mechanism not only provides a useful reference for processing on-site information but also enhances the accuracy of emergency decision-making.

(3)

The proposed method is primarily based on historical oil storage tank accidents, which may limit its generalizability to other industrial scenarios. In addition, uncertainties in multi-source data, the static nature of the constructed networks, and limitations of evidence theory may affect the accuracy of situational awareness. Future work will focus on extending the framework to other accident types, enhancing dynamic network updates, improving evidence theory-based fusion, and leveraging advanced technologies such as big data analytics and artificial intelligence to strengthen data integration.