Risk Management of Island Petrochemical Park: Accident Early Warning Model Based on Artificial Neural Network

Abstract

Island-type petrochemical parks have gradually become the ‘trend’ in establishing new parks because of the security advantages brought by their unique geographical locations. However, due to the frequent occurrence of natural disasters and difficulties in rescue in island-type parks, an early warning model is urgently needed to provide a basis for risk management. Previous research on early warning models of island-type parks seldom considered the particularity. In this study, the early warning indicator system is used as the input parameter to construct the early warning model of an island-type petrochemical park based on the back propagation (BP) neural network, and an actual island-type petrochemical park was used as a case to illustrate the model. Firstly, the safety influencing factors were screened by designing questionnaires and then an early warning indicator system was established. Secondly, particle swarm optimization (PSO) was introduced into the improved BP neural network to optimize the initial weights and thresholds of the neural network. A total of 30 groups of petrochemical park data were taken as samples—26 groups as training samples and 4 groups as test samples. Moreover, the safety status of the petrochemical park was set as the output parameter of the neural network. The comparative analysis shows that the optimized neural network is far superior to the unoptimized neural network in evaluation indicators. Finally, the Zhejiang Petrochemical Co., Ltd., park was used as a case to verify the accuracy of the proposed early warning model. Ultimately, the final output result was 0.8324, which indicates that the safety status of the case park was “safer”. The results show that the BP neural network introduced with PSO can effectively realize early warning, which is an effective model to realize the safety early warning of island-type petrochemical parks.

1. Introduction

The petrochemical industry is a pillar industry and the basic industry of the national economy [1], gathering various risk factors. Major accidents frequently occur in the petrochemical industry [2], and enterprises have to establish island-type petrochemical parks on islands close to the coast where there are few or no inhabitants to deal with social pressure [3]. For this reason, a safety early warning system [4] of an island petrochemical park has a significant role in the sustainable development of the park and in maintaining the stability of the production and society.

Many scholars have carried out research on early warning models [5,6,7], which are divided into four types, namely, single-indicator early warning model, multi-indicator early warning model, artificial intelligence early warning model, and comprehensive early warning model.

In a single-indicator warning model, several indicators are selected and detected to compare thresholds of each indicator and to judge a condition of the warning object. Although the single-indicator warning model is relatively intuitive, it can only act on descriptive analyses because of the lack of statistical and calculation tools. The single-indicator warning model often leads to the phenomenon that different prediction indicators are used to draw different conclusions for the same company. For this reason, the multi-indicator early-warning model gradually replaced the single-indicator one.

The multi-indicator models mainly select multiple indicators to give early warning to the research object at the same time. Kasprzyk-Hordern and Adams [8] analyzed potentially dangerous information by building a warning model with multiple indicators such as sewage, surface water, soil, and air. A multi-indicator linear model can reflect the operating status in multiple indicators, which easily reveal possible risks in all aspects and meet the requirements of enterprise warning [9]. However, a multi-indicator warning model can hardly achieve accurate judgment with missing data or low accuracy data. Therefore, despite multi-indicator warning models being more advantageous than single-indicator models, the intelligent early warning model has gradually become the ‘mainstream’ with the development of science and technology.

The intelligent early warning model is an early warning technology that uses intelligent technology and simulation technology to intelligently analyze the warning object, mainly including the artificial neural network (ANN) early warning model and support vector machine (SVM) early warning model. ANN relies on self-discovering rules without specifying, which is highly different from traditional methods [10,11]. Vinayagam Ramesh [12] predicted the adsorption effect by ANN, and the final results proved that the prediction effect was better. On the other hand, the SVM early warning model has massive advantages, such as simple structure, strong adaptability, global optimization, fast training speed, and strong generalization ability [13]. Bing et al. [14] used SVM to identify defects, and the finally result showed that the efficiency of identification based on SVM has greatly improved. The wide application of intelligent early warning shows its great potential in early warning research. Nevertheless, both ANN and SVM early warning model have enormous shortcomings in the data sample sizes [15,16]. Therefore, other scholars have begun to study the comprehensive early-warning model.

Multiple methods are combined for a comprehensive early warning model, which effectively makes up for the problem concerning a single early warning model easily getting into a local optimum and it hardly meets the applicable conditions. Zhou et al. [17] utilized the whale optimization algorithm (WOA) to optimize SVM and the model after optimizing has high accuracy. Li et al. [18] estimated the recidivism of criminals after being released by combination of a rough set and clustering algorithm, and the improved prediction result increased to 87.9%. Jiang et al. [19] combined the accelerated generalized autoregressive method and the Gauss–Cauchy mixed model to predict the value-at-risk of cryptocurrencies. Finally, compared with other prediction models, the model has a better prediction function. Zhang and Gao [20] improved the backpropagation (BP) neural network by the Grey Wolf optimization algorithm to give an early warning for the abnormal state of the electric vehicle charging voltage, and finally established a model with many effects for identifying the abnormality and early warning of the electric vehicle. Chi [21] predicted the electricity consumption in the region with wavelet transform and the LSTM model. The prediction error of the model was small, which proved the model had a good prediction function. Pavlicko et al. [22] calculated the maximum electricity consumption per day via various warning models, and finally found that the comprehensive warning model provided the best effects according to its performance results.

The island-type petrochemical park is built on an island without people or with few people; it has three characteristics: (1) highly corrosive: the surrounding environment of the island park is humid, and the salt spray will accelerate the aging and corrosion of the equipment [23,24]. Thus, this environment easily causes the leakage of toxic gases, such as flammable and explosive, which increases the risk of major accidents [25]. (2) Extreme weather: the island-type petrochemical parks are vulnerable to extreme weather, such as typhoons, heavy rainfall, and lightning strikes, etc. When severe weather synthesizes, the consequences become more serious. (3) Weak rescue force: the island-type park is geographically closed, with poor traffic on the island, and relatively simple access channels [26]. This location condition makes the external emergency and rescue materials hardly arrive on time when a dangerous situation or serious accident occurs. However, the current research on early warning models of petrochemical parks is not sufficient, and the only research rarely takes into account the particularities brought about by the geographical locations of islands; that is, strong corrosion, extreme weather, and weak external rescue forces. The established early warning models are not suitable for an island-type petrochemical park.

This paper comprehensively considers the influencing factors through an accident causative factor analysis, field research, and expert consultation. Further, the indicator system was constructed through a questionnaire analysis and a safety warning model based on the BP neural network was constructed with the indicator system as input parameters. Although the BP neural network has strong self-learning, self-adapting ability, and nonlinear mapping ability, the BP network still easily drops into a local optimum and needs numerous parameters to complete training [27,28]. Therefore, particle swarm optimization (PSO) is used to optimize the BP neural network to improve its convergence rate and global optimization ability. This study improves the safety level of island petrochemical parks, realizes timely prevention, and early warning of accidents, maintains social peace and stability, promotes the improvement of quality of life, and promotes the process of urbanization.

2. Construction of the Safety Early Warning Indicator System

The petrochemical park has many characteristics, such as a complex production process, close connection, a huge scale, and so on. In order to achieve the expected effect of the early warning indicator system, the system demands to complete and correspond to the actual situation. Therefore, this paper is based on the analysis of causative factors in chemical accidents to obtain influencing factors, which compose the safety early warning indicator system. In addition, the methods of on-site investigation, consultation with experts, and analysis of existing historical documents are combined to completely analyze the relevant factors. The safety factors of island petrochemical parks are divided into four categories in this study—people, objects, environment, and management. The breakdown of each type of factor below is shown in

Table 1. Factors affecting the safety of island petrochemical parks.

First-Level Indicator	Second-Level Indicator
People	Operation proficiency
	Illegal operation
	Personnel physical condition
	Personnel psychological situation
	Personnel holding certificates
	Physical examination pass rate
	Average working years
	Safety assessment pass rate
Objects	Reasonable material selection
	Material tool stacking rationality
	Completeness of safety guards
	Equipment failure rate
	Completeness of protective equipment
	Monitoring and testing facility integrity rate
	Incompleteness of fire extinguishing facilities
	Degree of damage to the equipment
Environment	Park traffic conditions
	Layout of the park site
	Climate conditions
	Geological conditions
	Anti-noise and dust measures
Management	The rationality of allocation of personnel
	Perfect emergency plan rate
	Emergency drill implementation rate
	Safety production inspection
	Safety investment
	Safety education
	Hazard inspection and rectification
	Improvement of safety management system
	Effectiveness of government oversight
	Occupational hazard protection effectiveness

.

2.1. Rationality of Indicator System

In order to judge the rationality of the indicator system, it is necessary to conduct a questionnaire survey on the importance of the indicator system and influencing factors. The questionnaire is in the form of the Likert five-level scale [29,30]; the level means that the degree of influence gradually increases. There are a total of 100 questionnaires distributed and 81 valid questionnaires were recovered. The Cronbach’s alpha coefficient [31], Kaiser–Meyer–Olkin (KMO) [32], and Bartlett’s test of sphericity were used to test the reliability and validity of the questionnaire [33,34] in this study. Those methods also judged the degree of consistency with the safety factors of the 31 island-type petrochemical parks listed in Table 1. Typically, the data of the questionnaire are believed to be reasonable when both the Cronbach’s alpha coefficient and the KMO are greater than 0.7.

2.1.1. Reliability Test

The reliability test is a standard to measure the reliability of questionnaire data. This paper adopts the commonly used Cronbach’s alpha coefficient method; the lager the coefficient, the higher the reliability of the questionnaire. Typically, the coefficient ranges from 0 to 1; the specific value standard is shown in

Table 2. Value range of Cronbach’s α coefficient.

Range of Coefficient	Guideline
0.9 ≤ α	Very high reliability, believable
0.8 ≤ α < 0.9	High reliability, credible
0.7 ≤ α < 0.8	Reliability, credible
0.6 ≤ α < 0.7	Certain problems in the questionnaire, which has reference value
α < 0.6	Credibility questionnaire is very problematic and needs to be redesigned

. The Cronbach’s α coefficient is 0.922, by testing 31 variables via SPSS software trial version (IBM, Armonk, NY, USA), and is greater than 0.9, indicating that the data obtained in this file conform to the needs of the statistical analysis.

2.1.2. Validity Test

Validity refers to the accuracy in which the questionnaire results correctly reflect the characteristics of the questionnaire objects. The higher the validity, the more effective the questionnaire is. Thus, KMO and Bartlett’s test of sphericity were used to verify the correctness. The KMO inspection standard is shown in

Table 3. KMO inspection criteria.

KMO	Evaluation Criterion
0.9 ≤ KMO	Very suitable for factor analysis
0.8 ≤ KMO < 0.9	Suitable for factor analysis
0.7 ≤ KMO < 0.8	Fit for factor analysis
0.6 ≤ KMO < 0.7	Can be done for factor analysis
0.5 ≤ KMO < 0.6	Barely suitable for factor analysis
KMO < 0.5	Unsuitable for factor analysis

. KMO and Bartlett’s test of sphericity were performed on 31 factors with the help of SPSS software, as shown in

Table 4. KMO and Bartlett test.

Adequacy sample of KMO metric		0.802
Bartlett’s Test of Sphericity	Approximate chi-square	1305.221
	dF	465
	Sig	0.000

. As can be seen, the KMO test value in the software analysis is 0.802, which indicates that the correlation between variables has little discrepancy. The accompanying probability value of the sphericity test is 0.000, which is less than the significance level of 0.001, indicating that the validity meets the analysis requirements.

2.1.3. Indicator Screening

Among the 31 influencing factors in the questionnaire, there may be a linear relationship between the factors in terms of their degrees of influence, which means a certain factor can be represented by a linear equation of one or several other factors. However, this linear relationship easily makes an information overlap when using these factors to construct an indicator system. Similarly, some factors that have little impact on establishing the warning indicator system can be discarded. In this paper, the factor analysis method is used to achieve dimensional reduction and deal with the linear relationship between the influencing factors or the degree of influence. This method makes the selection of early warning indicators more scientific and converts 31 factors into a few more that are more representative. According to the extraction principle of the cumulative contribution rate of variance n ≥ 90%, the explained variance table was calculated with the help of SPSS software, and a total of 18 comprehensive factors were extracted. The cumulative contribution rate of variance is 90.021%. This suggests that the extracted 18 comprehensive factors reflect 90.021% of the characteristic information of 31 influencing factors.

2.2. Establishment of the Early Warning Indicator System

As described, the finally established island-type petrochemical park safety early warning indicator system consisted of 4 first-level indicators and 18 second-level indicators in this paper, as shown in Figure 1.

3. Safety Early Warning Model of Island-Type Petrochemical Park

3.1. BP Neural Network

BP neural network is a multi-layer forward neural network model with information forward propagation and error backpropagation, which contains single or more hidden layers [4,28]. Further, this type of layer enables performing nonlinear mapping, between inputs and outputs, and solving linear inseparable problems. With the advantages of parallel processing, distributed storage, and fault tolerance in structures, the BP neural network has the characteristics of self-learning, self-organization, and self-adaption [16,27]. The schematic diagram of the BP neural network structure is shown in Figure 2.

The PSO algorithm was introduced to solve the shortcomings in the parameter settings of the traditional BP network. Each individual in the PSO algorithm was regarded as a particle without weight and volume in the n-dimensions search space, which has a certain speed in the space [35]. In particular, the flight speed of the particle was dynamically adjusted by the experience of the individual flight and group flight. A total of 18 early warning indicators were taken as input nodes and the overall safety status was taken as an output node of the BP neural network.

3.2. Model Prediction Process

The process of the safety early warning model for the island-type petrochemical park based on the PSO-BP neural network is shown in Figure 3. At first, PSO was used to optimize the initial weights and thresholds of the BP neural network, then the obtained optimal solutions were set as the new weights and thresholds. Finally, the fully trained network was used to predict and simulate the safety status of the island-type petrochemical park.

4. Case Study

The data from 30 petrochemical parks were set as training samples to train the proposed PSO-BP early warning model of island-type petrochemical parks, meanwhile the Zhejiang Petrochemical Co., Ltd., in Zhoushan City, Zhejiang province, was selected as a case analysis to verify the accuracy of the proposed warning model.

4.1. Parameter Setting

4.1.1. Parameter Setting of BP Network

The focus of this paper was to establish an index system and a PSO-BP petrochemical park early warning model based on the index system. In this study, the sample data were gathered form the arithmetic average of the indicator data and the overall security status of 30 petrochemical parks by 10 experts. Moreover, 26 of 30 sample data were training samples and the remaining 4 samples were test samples to verify the results. In principle, the BP neural network needs more data, and the more data, the better the accuracy of the model. We did our best to obtain 30 precious sets of data; however, it was difficult to collect thousands of instances. In order to make up for the shortage of the small amount of data, we introduced the PSO algorithm to optimize the BP neural network. To facilitate the readers of the paper in finding innovative data in this work, the 30 complete datasets that allowed replication of the results were collated and uploaded to the submission system as Supplementary Materials.

(1)

Input parameters.

A total of 18 indicators in Section 2.2, of the established island-type petrochemical park safety early warning indicator system, were taken as input parameters; the specific scoring standards are shown in

Table 5. Input indicator scoring criteria.

Score	Grading System
90–100	Very safe
80–89	Safer
70–79	Safe
60–69	Less safe
0–59	Unsafe

.

(2)

Output parameter.

In this study, the safety warning indicator represents the overall safety status of the petrochemical park, as shown in

Table 6. Security rating scale.

Safety Level	Score
Very safe	[0.9, 1]
Safer	[0.8, 0.9)
Safe	[0.7, 0.8)
More dangerous	[0.6, 0.7)
Very dangerous	[0, 0.6)

. Therefore, the safety level that was used as the output parameter of the BP neural network.

(3)

Normalization.

The normalization was used to eliminate the impact of different field weights on the model performance, which commonly used data processing methods, including the minimax method and the mean-variance method. Thus, the minimum and maximum value method was selected to normalize the data to interval [0, 1], applying the map minmax function provided by MATLAB trial version (MathWorks, Natick, MA, USA). A milestone on the normalization and backpropagation work came from Lecun et al. [36]; they analyzed the convergence of backpropagation learning, provided tips to avoid many of the bad behaviors of backpropagation, and explained why they worked. It is recommended that the readers study it carefully so that they can better design the backpropagation and set the parameters required by the training network.

(4)

Layer design of the BP neural network.

There are input layers, output layers, and hidden layers in the BP neural network that the number of nodes in each layer need to determine. The number of input and output nodes were, respectively, determined to be 18 and 1, according to the constructed indicator system.

Although a conventionally single hidden layer is enough to complete the nonlinear mapping, the hidden layer can also be set as multiple layers. Ultimately, this study sets a single hidden layer. As for the number of nodes in the hidden layer, there is no other good method at present, most of which refer to historical experience. However, on the basis of referring to relevant research, the prediction effect is the most ideal when the number of nodes in the hidden layer is 12.

(5)

Training method.

The choice of training method is related to the risk prediction problem and the number of training samples. The Levenberg–Marquardt algorithm [37,38], with fast convergence speed and a small mean square error, was chosen as the training method of the predict model in this paper. In addition, the sigmoid activation function was used, and the loss function was the mean square error function.

4.1.2. Parameter Setting of PSO

Typically, the parameter settings will directly affect the optimization performance of the algorithm when PSO is used to optimize the parameters. The main parameters involved in PSO are acceleration factors c_i, maximum speed Vmax, population size N, and inertia weight w, as shown in

Table 7. Parameter values.

Parameters of PSO	Value
Acceleration factors ci	c1 and c2 both are 1.49445
Population size N	10
Maximum speed Vmax	1
Inertia weight w	0.8

.

4.1.3. Model Evaluation Indicators

The model evaluation indicators are used to judge the quality of the safety early warning model, which include the error sum (ES), the mean absolute percentage error (MAPE), the root mean square error (RMES), and the mean absolute error (MAE).

(1)

Error sum (ES).

$$\begin{array}{c}\mathrm{E}\mathrm{S}={\displaystyle {\displaystyle \sum }_{i=1}^N}\left|x_i-{\widehat{x}}_i\right|\end{array}$$

(1)

where $x_i$ represents the measured value of the sample, ${\widehat{x}}_i$ means the prediction value, $N$ represents the number of samples.

(2)

Mean absolute percentage error (MAPE).

$$\begin{array}{c}\mathrm{M}\mathrm{A}\mathrm{P}\mathrm{E}=\frac{1}{N}{\displaystyle {\displaystyle \sum }_{i=1}^N}\left|\frac{x_i-{\widehat{x}}_i}{x_i}\right|\times 100\%\end{array}$$

(2)

where the symbols in Equation (2) have the same meaning as in Equation (1).

(3)

Root mean square error (RMSE).

$$\begin{array}{c}\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{{{\displaystyle \sum }}_{i=1}^N{\left(x_i-{\widehat{x}}_i\right)}^2}{N}}\end{array}$$

(3)

where the symbols in Equation (3) have the same meaning as in Equation (1).

(4)

Mean absolute error (MAE).

$$\begin{array}{c}\mathrm{M}\mathrm{A}\mathrm{E}=\frac{1}{N}{\displaystyle \sum }_{i=1}^N\left|x_i-{\widehat{x}}_i\right|\end{array}$$

(4)

where the symbols in Equation (4) have the same meaning as in Equation (1).

4.2. Sample Training

This study used MATLAB software to implement the early warning model based on the BP neural network improved by the PSO algorithm. A total of 18 indicators (in Section 2.2) were taken as input parameters, while the number of neuron nodes corresponding to the input layer was also 18, and the safety status of the petrochemical park was used as an output parameter, one hidden layer with 12 nodes was also selected, as shown in Figure 4. A total of 26 of 30 sample data were training sets and the remaining 4 were test sets to verify the model accuracy. The error reached the design accuracy requirement of 0.00001 after the program was trained four times, and the early warning improved model converges.

As shown in

Table 8. Error comparison of evaluation indicators of the 26 training sets.

	ES	MAPE	RMSE	MAE
BP neural network	0.15772	4.6368	0.041618	0.039431
PSO-BP neural network	0.10007	2.9604	0.031138	0.025017

, all kinds of errors for 26 training sets after optimization (PSO-BP) were smaller than the network without optimization (BP), which indicates that the PSO algorithm has an improved effect on the BP neural network. In addition, as shown in

Table 9. Comparison of the predicted values of the four test sets.

Test Sample	1	2	3	4
Sample true value	0.85	0.83	0.82	0.84
BP test results	0.8708	0.8131	0.8242	0.8758
PSO-BP test results	0.8632	0.8271	0.8230	0.8411

, the result predicted by the PSO-BP neural network was closer to the true value and more accurate than the result predicted by traditional BP network. From Table 8 and Table 9, it can be seen that the PSO algorithm has an improved effect on BP, and PSO-BP can be used for the prediction of the early warning model.

4.3. Model Application

Zhejiang Petrochemical Co., Ltd., is located in Yushan Island, Daishan County, Zhoushan City, Zhejiang province. It produces more than 20 kinds of petrochemical products, such as gasoline and diesel, jet fuel, paraxylene (PX), high-end polyolefin, and polycarbonate. Facilities and equipment, such as crude oil terminal, storage, transportation, oil refining/chemical production, processing, product outbound transportation and service facilities, personnel, and materials, can be transported through the Yushan Bridge on the island and the self-built wharf. The geographic location is shown in Figure 5.

This study collected indicator data of the project by an expert group scoring method and the arithmetic mean scoring is the final score of the indicator in the system. The final score of each indicator is 0.87, 0.85, 0.75, 0.82, 0.83, 0.83, 0.84, 0.84, 0.83, 0.85, 0.864, 0.888, 0.876, 0.8, 0.82, 0.83, 0.8, 0.85, respectively. Ultimately, the above data were substituted as input parameters into the PSO-BP early warning model for park safety assessment. The output result was “test_sim = 0.8324”, which represents the safety rating of “Zhejiang Petrochemical Co., Ltd.”; it was 83.24, which is in the range of “Safer”. This result is consistent with the actual situation troughing the investigation of the daily production operation of “Zhejiang Petrochemical Co., Ltd.”. Therefore, the early warning model of the island petrochemical park established in this paper is feasible in practice.

5. Conclusions

A total of 31 influencing factors were obtained by analyzing the hazard factors of the recorded historical petrochemical park chemical accidents. Based on the factors obtained from the analysis, a questionnaire was designed to screen the influencing factors to avoid the interference of the repeated information on the early warning system model. It can be concluded that the reliability and validity of the questionnaire that was analyzed by software SPSS was within the standard range. A total of 4 first-level indicators and 18 second-level indicators were obtained after rearranging and grading 31 influencing factors.

Although the threshold and initial weight in the early warning model determined the effectiveness and safety of the entire model, both of them will not be optimized and iterated in the traditional BP neural network. Therefore, the above parameters were optimized by the PSO algorithm in this study. Each particle in PSO was used to respectively correspond to the threshold and initial weight. By constantly comparing the fitness of different positions, one must find the optimal value as the threshold and the initial weight before the data enters the hidden layer of the BP neural network.

The arithmetic means of the evaluation and scoring of 30 typical petrochemical parks by the expert group were used as the final sample data. In addition, all 30 sets of samples were divided into training sets and test sets. The scores evaluated by 18 indicators were used as input indicators, and the output indicators were the overall security status of the park. To summarize, the case study of Zhejiang Petrochemical was selected to carry out the case verification of the island-type petrochemical park, which confirms the reliability of the model. In the future, the proposed methodology will be supplemented and improved by collecting more data to provide more insightful implications and improve the accuracy of the model. Nevertheless, this paper realizes early warning of island petrochemical parks based on the indicator analysis method and PSO-BP model, which has good reference value and provides a basis for supervision by relevant government departments.