You are on page 1of 8

Journal of Loss Prevention in the Process Industries 63 (2020) 103988

Contents lists available at ScienceDirect

Journal of Loss Prevention in the Process Industries


journal homepage: http://www.elsevier.com/locate/jlp

Fused CFD-interpolation model for real-time prediction of hazardous gas


dispersion in emergency rescue
Yan Jiang a, b, Zhengyi Xu b, Jianming Wei b, Guowei Teng a, *
a
School of Communication and Information Engineering, Shanghai University, Shanghai 200444, China
b
Shanghai Advanced Research Institute, Chinese Academy of Sciences, Shanghai 201210, China

A R T I C L E I N F O A B S T R A C T

Keywords: Accidental releases of hazardous gases in chemical industries can pose great threats to public security. The
Hazardous gas dispersion computational fluid dynamics (CFD) model is commonly applied to predict gas dispersion in complex structured
Emergency response areas. It can provide good accuracy but it is too time-consuming to be used in emergency response. To reduce
Real-time prediction
computation time while keep acceptable accuracy, this paper proposes several fused CFD-interpolation models
Computational fluid dynamics
which combine CFD model with different interpolation methods. Spline, linear and nearest interpolation
methods are used. A CFD simulations database is created ahead of time which can be quickly recalled for
emergency usage and unknown situations can be predicted instantly by interpolation methods instead of time-
consuming CFD model. Fused models were applied to a case study involving a hypothetical propane release
with varying conditions and validated against CFD model. The validation shows that prediction accuracy of these
fusion models is acceptable. Among these models, CFD-Spline interpolation model performs best. It is faster than
CFD model by a factor of 75 and is potentially a good method to be applied to real-time prediction.

1. Introduction 2018), computational fluid dynamics (CFD) model (de OB Lira et al.,
2018; Kakosimos and Assael, 2013; Liu et al., 2015; Na et al., 2018) and
In chemical industrial parks, there are a large number of chemicals, so on. For different gas dispersion models, their applicable scenarios,
many of which are flammable, explosive and hazardous gases. Usually, if calculation accuracy and calculation speed are all different. Table 1
the relative density of a gas is larger than that of air, we call it heavy gas summarizes the advantages and limitations of Gaussian model and CFD
(Dong et al., 2017; Mack and Spruijt, 2013). Most of the flammable or model. In an emergency scenario, a best model should be fast and ac­
hazardous gases are heavy gases. Accidental release of heavy gases curate. However, this perfect model does not exist yet.
forms a low momentum continuous release source or instantaneous Attempts were made to combine these two capabilities. Wang et al.
source near the ground. Due to its own gravity, heavy gas has a large (2015) developed a fast prediction approach by combining neural net­
lateral pollution distance after leakage which is easy to cause extremely works with the PHAST model. Trained neural networks could obtain
serious harm to humans in a short time. To minimize harm to humans acceptable results in real time. But this method was only implemented in
and avoid further accidents, it is of great significance to quickly and a 2-D domain and could not be applied to the actual situation. Then,
accurately predict heavy gas dispersion in complex structured chemical Wang and Qian (2018) extended his existing work from 2D to 3D by
industrial parks. Using the gas dispersion model to obtain the distribu­ combining cellular automata and artificial neural networks. The cellular
tion of gas concentration is the basis for quickly and accurately automata model could calculate the gas dispersion process about 1.5
designing the escape route, determining the accident evacuation zone times faster than the fire dynamic simulator. However, the calculation
and formulating a public early warning plan and other emergency rescue time was still too long to meet the needs of emergency rescue. Ma and
assistance decision-making. Zhang (2016) combined machine learning methods with Gaussian
Currently, there are several models that can simulate the dispersion models to improve the prediction accuracy of machine learning pre­
process after the leakage of hazardous gases, such as Gaussian dispersion diction networks. However, in the prediction results, the concentrations
models (Askariyeh et al., 2017; Gibson et al., 2013; Matacchiera et al., at the locations far from the leak source were overestimated and even

* Corresponding author.
E-mail address: tenggw@shu.edu.cn (G. Teng).

https://doi.org/10.1016/j.jlp.2019.103988
Received 2 March 2019; Received in revised form 2 October 2019; Accepted 22 October 2019
Available online 30 October 2019
0950-4230/© 2019 Elsevier Ltd. All rights reserved.
Y. Jiang et al. Journal of Loss Prevention in the Process Industries 63 (2020) 103988

Table 1 b. CFD model is used to simulate gas dispersion under different leak
Summary of existing models for gas dispersion estimation including their ad­ scenarios in the target area. Since numerous parameters impact gas
vantages and limitations. dispersion in the atmosphere and are required to accurately simulate
Model Advantages Limitations Sources the dispersion, a general leak scenario contains several parts: (1)
Gaussian Relatively simple and Inadequate for Askariyeh et al.
source information: the release material, source location and source
model easy to use; fast complex structured (2017); Gibson et al. strength; (2) weather condition: wind speed, wind direction and
computation; areas. (2013); Matacchiera temperature. By varying the value of some or all of these parameters,
generate dispersion et al. (2018); Na different leak scenarios are generated. The gas concentrations ob­
distances which are et al. (2018)
tained from CFD simulations are stored in a database. It is worth
perceived as being
conservative; noticing that Step a and b should be performed in advance in a real-
effective in flat and time prediction of gas dispersion.
unobstructed terrain. c. When a gas leakage accident occurs, get information including the
CFD Consider the effects of Time-consuming; Kakosimos and release material, release velocity, wind speed and wind direction.
model terrain and obstacles; not suitable for Assael (2013);
suitable for accident emergency Mishra et al. (2014);
d. CFD model is used to simulate the target leakage accident. The ge­
analysis and response; need Kurnia et al. (2014); ometry model needs to be meshed before simulation. The simulation
parametric studies. professional users Liu et al. (2015); Li results of the CFD model include: the position coordinates of each
et al. (2016); Liu grid point in the target area, and the gas concentration value at each
et al. (2016);
coordinate. The results of the CFD model will be compared with the
Dasgotra et al.
(2018); de OB Lira results of the interpolation method to evaluate the performance of
et al. (2018); Rum the interpolation method.
et al. (2018) e. Interpolation methods are used to predict gas concentrations of the
target leakage accident. The CFD database generated in Step a is used
as the input of the interpolation methods. The output of interpolation
worse than the results of the original Gaussian model. Lauret et al.
methods is the gas concentrations of the situation to be predicted.
(2016) created a large database of CFD calculations and combined
f. The performance of the prediction model is evaluated by comparing
neural networks and cellular automata to predict gas dispersion. This
the prediction results of the model with the data generated by CFD
method can obtain better results on high wind velocities cases but worse
model.
results on low wind velocities cases. Further, Lauret et al. (2017) have
used artificial neural network (ANN) to forecast gas concentrations
without knowing the information of leakage source, but this model 2.2. Configuration of the fused CFD-interpolation model
tended to underestimate high concentrations and overestimate low
concentrations. 2.2.1. Interpolation method
This paper focuses on the real-time prediction of gas leakage in The CFD model is complex and the gas concentration is related to
complex environments in emergency response. To solve the problem of multiple parameters. In order to simplify the prediction model to in­
long calculation time of CFD model, this paper proposes several CFD- crease its calculation speed, only a few parameters that have a large
interpolation models which combine CFD model with different inter­ influence on gas concentrations are considered. Wind speed, wind di­
polation methods. In these models, CFD is used to simulate numbers of rection and gas release rate are three main factors affecting gas con­
gas leakage scenarios that could cover almost all situations. Thus, a large centrations (Wang et al., 2013). The interpolation method is used to
database of CFD calculations is created. When a gas leak accident occurs, correlate the gas concentrations to the above three parameters, and
the accident information will be matched with the information in the simplifies the prediction model.
database. If the information of accident just exists in the database, the The way of interpolation is to find the approximate function of a
predicted gas concentrations can be quickly obtained directly from the discrete set of known data points. By using the interpolation method, the
database. If the accident is an unknown situation not included in the value of some unknown points can be predicted with only a limited
database, the linear, nearest or spline interpolation method will be used number of data points. The concentrations to be predicted can be ob­
to interpolate the data in the database. As a result, gas concentrations in tained by simple interpolation calculation without solving the complex
a leakage accident can be obtained in a short time. The CFD- Navier-Stokes equation. The time of gas concentration prediction is
interpolation models will be verified by using a case study of hypo­ reduced in this way.
thetical propane release. There are some commonly used interpolation methods such as linear
This paper is divided into the five parts: after the introduction, part 2 interpolation, cubic spline interpolation and nearest interpolation. In
details the fused CFD-interpolation model proposed in this paper. A this paper, the prediction model of contaminant dispersion was
hypothetical propane release case and prediction results obtained using respectively built on linear interpolation, nearest interpolation, and
the fused model are shown in part 3 and 4, respectively. A conclusion is cubic spline interpolation method. The prediction performance of
proposed in part 5. different methods was compared.

2. Methodology 2.2.2. Linear interpolation


Linear interpolation is a method of joining points by straight lines.
2.1. Workflow of the fused CFD-interpolation model Each line is bounded by two adjacent data points and can be interpolated
independently. Suppose ðxi ; yi Þ is the previous endpoint of a line segment
In this paper, it is assumed that the result of the CFD model is the and ðxiþ1 ; yiþ1 Þ is the latter endpoint of the line segment. For the point
actual situation of gas leakage. The proposed methodology to develop where the abscissa is x and x is in the range ½xi ; xiþ1 �, its y is
the fused model for predicting real-time gas dispersion contains 6 steps, x xi
y ¼ yi þ ðyiþ1 yi Þ (1)
as shown in Fig. 1. xiþ1 xi
When input parameters are multidimensional (as described in Sec­
a. First, collect environmental information including plant layout of the
tion 2.2.5), bilinear interpolation or trilinear interpolation is required.
target area. Then, build a 3D model of the target area.
Bilinear interpolation is a linear interpolation with two variables. The
main idea is to perform interpolation calculation (Eq. (1)) respectively in

2
Y. Jiang et al. Journal of Loss Prevention in the Process Industries 63 (2020) 103988

Fig. 1. Flowchart of the integrated CFD-interpolation model.

directions of the two variables. What’s more, the trilinear interpolation wise polynomial called spline. Suppose there are n þ 1 nodes ðx0 ; x1 ; ⋯
is to add one direction (or variable) based on the bilinear interpolation ; xn Þ on the interval ½a; b�. The value yi ði ¼ 0; 1; 2; ⋯; nÞ corresponding to
(Bourke, 1999). each node is known. The cubic spline interpolation is to construct a
cubic polynomial in each cell ½xj 1 ; xj �ðj ¼ 1; 2; ⋯; nÞ. The cubic spline
2.2.3. Nearest interpolation equation must satisfy the following conditions:
The nearest interpolation is the simplest interpolation method. Its
main idea is to find a point closest to the interpolation point and then 1. In each cell ½xi 1 ;xi �ði ¼ 1; 2;⋯;nÞ, SðxÞ ¼ Sðxi Þ is a Cubic polynomial,
assign the value of the point to the interpolation point. 8
When input parameters are multidimensional (as described in Sec­ < C1 ðxÞ; x0 � x � x1
SðxÞ ¼ Ci ðxÞ; xi 1 � x � xi (2)
tion 2.2.5), the idea of multidimensional nearest neighbor interpolation :
Cn ðxÞ; xn 1 � x � xn
is similar to one-dimension nearest interpolation. The nearest interpo­
lation is extended by applying it independently to each dimension.
where Ci ðxÞ represents a cubic polynomial function for x, its function
is
2.2.4. Cubic spline interpolation
Spline interpolation is a form of interpolation using a special piece­

3
Y. Jiang et al. Journal of Loss Prevention in the Process Industries 63 (2020) 103988

Ci ðxÞ ¼ ai x3 þ bi x2 þ ci x þ di (3) � �2 �
VG ¼ exp lnCo lnCp (7)


2. Sðxi Þ ¼ yi ði ¼ 1; 2; ⋯; nÞ. ðCo Co Þ Cp Cp
R¼ (8)
3. SðxÞ, S ðxÞ and S00 ðxÞ are continuous on the interval ½a; b�. S ðxÞ is the
0 0
σ Cp σ Co
derivative of SðxÞ. S ðxÞ is the second derivative of SðxÞ.
00

Cp
FAC2 ¼ fraction of data that satisfy 0:5 � � 2:0 (9)
Therefore, as long as ai , bi , ci and di are determined, the interpolation Co
function SðxÞ is determined. Similar to bilinear interpolation and
trilinear interpolation, when input parameters are multidimensional (as where Cp represents predicted concentration, Co represents observed
described in Section 2.2.5), we can extend the spline interpolation concentration, C is the average value over the concentration dataset and
method to a higher dimension by applying it independently to each σ C is the standard deviation over the dataset.
dimension (Bourke, 1999). For an ideal model, the value of MG, VG, R and FAC2 should be 1 and
the value of FB and NMSE should be 0. However, such a model does not
2.2.5. Input and output strategies exist. The prediction model is defined as acceptable when the values of
The input parameters of the interpolation function should be readily the indicators are within the following ranges (Chang and Hanna, 2004):
available at the scene of an accident. Three main factors affecting gas 0:3 < FB < 0:3, 0:7 < MG < 1:3, NMSE < 4, VG < 1:6, FAC2 > 0:5.
dispersion are wind speed, wind direction and gas release rate (Wang
et al., 2013). Environmental parameters such as wind speed and direc­ 3. Numerical case study
tion can be obtained directly by gas sensors, and gas release rate need to
be obtained by inversion algorithms (Ma et al., 2017; Zheng and Chen, A hypothetical propane release case was introduced to test this
2011). The inversion algorithm is not discussed in this paper and these model. The geometric model establishment, example database creation
input parameters are assumed to be already known. and interpolation of the prediction model were performed using the
Different wind speeds (X1 ), wind directions (X2 ) and release rate (X3 ) previous guidelines.
make up different leakage scenarios. Gas concentrations X4 in different
scenarios (X1 , X2 , X3 ) are also the inputs of the interpolation function. 3.1. The geometric model of the target area
Actually, X4 is the gas concentration data in the CFD database
mentioned in Section 2.1. In the case where (X1 , X2 , X3 , X4 ) are known, First, a 3D model of the target area was built. The target area was
what we have to do is to predict the gas concentrations (Y) under a new defined as a 30 m long, 20 m wide and 5 m high domain (Fig. 3). There
situation (XI1 , XI2 , XI3 ). Therefore, the input parameters of the proposed are two buildings in Fig. 3. The building on the left serves as the source
method are XI1 - XI3 . The output parameter is the unknown Y. The input of leakage. Assume that the gas releases from a circular hole with a
and output parameters of the interpolation method are shown in Fig. 2. radius of 100 mm at the top of the leakage source. The other one serves
as an obstacle. The boundary conditions are also shown in Fig. 3. As
2.3. Performance criteria shown in Fig. 3, the boundaries (X ¼ 0, Y ¼ 0) are velocity-inlet. The
flow direction of velocity-inlet1 is defined as the positive x-axis direction
To evaluate the performance of the air quality prediction model, and the flow direction of velocity-inlet2 is defined as the positive y-axis
several performance criteria including the fractional bias (FB), the direction. The boundaries (X ¼ 30, Y ¼ 20) are outflow. Other bound­
geometric mean bias (MG), the normalized mean square error (NMSE), aries not mentioned are all solid walls.
the geometric variance (VG), the correlation coefficient (R) and the A schematic diagram of three-dimensional coordinate system is
fraction of predictions within a factor of two of observations (FAC2) added in the lower left corner of Fig. 3. In the simulation domain, the
were proposed by Chang and Hanna (2004). All of these criteria were 0� wind direction is defined as the positive x-axis direction and the 90�
used in this paper. These criteria are defined as follows (Chang and wind direction is defined as the positive direction of y-axis.
Hanna, 2004): The following examples are used to illustrate the specific settings of
wind direction and wind speed:
Co Cp
FB ¼ � (4) First, when the wind direction is 0o and the wind speed is 2 m/s, the
0:5 Co þ Cp
velocity magnitude of velocity-inlet1 is 2 m/s, and the velocity magni­
� tude of velocity-inlet2 is 0.
MG ¼ exp lnCo lnCp (5)
Second, when the wind direction is 90o and the wind speed is 2 m/s,
�2 the velocity magnitude of velocity-inlet1 is 0, and the velocity magni­
Co Cp tude of velocity-inlet2 is 2 m/s.
NMSE ¼ (6)
Co Cp Third, as shown in Fig. 4, when the wind direction is 30o and the
wind speed is 2 m/s, the velocity magnitude of velocity-inlet1 is 1.73 m/

Fig. 2. Input and output parameters of interpolation methods. Fig. 3. Sketch of the geometric model of the target area.

4
Y. Jiang et al. Journal of Loss Prevention in the Process Industries 63 (2020) 103988

s, and the velocity magnitude of velocity-inlet2 is 1 m/s. Table 2


Test cases characteristics.
3.2. Example database creation Case Id Wind speed (ms 1
) Wind direction(� ) Release rate (g=s)

1 2.1 33.5 65
The geometry model needs to be meshed before simulation. It should 2 2.8 33.5 77
be noted here that the concentrations we finally obtained were the 3 3.7 32.1 73
concentrations on the horizontal section at a height of 1.6 m. The con­ 4 4.6 34.5 83
centration at 1.6 m was chosen for analysis mainly because it was 5 5.6 38.5 86
6 6.9 38.5 94
considered that people’s height was mostly around 1.6 m. Then, CFD
model was used to simulate gas dispersion under different conditions.
Here, the classic RANS (Reynolds-Averaged Navier-Stokes) k-ε model compared.
was used. It has to be reminded that, in this paper, regardless of the Fig. 5 shows the Cp and Co of test case 1, black lines represent Co and
accuracy of the CFD model, the simulation results of CFD model is red lines represent Cp, but we can’t see the difference clearly between
assumed to be the actual results of gas dispersion. them at all. Enlarge part of Fig. 5 to get Fig. 6.
By varying the values of wind speed, wind direction and gas release We still can’t see the difference between Cp and Co, which means
rate, 605 simulations were performed, and the gas concentrations ob­ that the difference between Cp and Co is very small. In order to see the
tained for each simulation were stored in a database. difference between Cp and Co intuitively and compare the prediction
Specifically, all the 605 simulations can be decomposed into: accuracy of different models, we need to rely on those performance
criteria. As can be seen from Equations (4)–(9), the performance criteria
� 11 variations of wind direction varying from 30o to 40o with a 1o (NMSE, FB, VG, etc.) are essentially the difference between Cp and Co.
interval; From the data in Figs. 7–9 and Table 4, we can clearly see the difference
� 11 variations of wind speed varying from 2 m/s to 7 m/s with a in accuracy between different methods.
0.5 m/s interval; The prediction performance of different models is analyzed by per­
� 5 variations of gas release rate varying from 60 g/s to 100 g/s with a formance criteria proposed by Chang and Hanna (2004). The perfor­
10 g/s interval. mance criteria for an acceptable predictive model are shown in Table 3.
For these performance criteria especially FB and NMSE are sensitive
Each one of the 605 simulations is composed of 3,718 concentra­ to extreme high and low concentrations, it is necessary to set a minimum
tions. These values correspond to 3,718 coordinate points on the hori­ threshold for observed and predicted concentrations (Dong et al., 2017).
zontal section at a height of 1.6 m in target area. That is to say, in these The minimum threshold in this study was 10 5 .
605 cases, each coordinate has 605 concentration values. By interpo­ Figs. 7–9 show the values of NMSE, FB and MG obtained by using
lating the values of each point and calculating the concentration values three fused models in case 1–6, respectively. Since the values of NMSE
under certain conditions, 3,718 predicted values can be obtained, and a are relatively small, and the magnitude difference is somewhat large, the
complete gas dispersion profile can be formed. values of and log10 (NMSE) are taken for Fig. 7.
According to Section 2.3, the closer the values of FB and NMSE are to
3.3. Case parameter settings 0, the smaller the error of the prediction model is. Meanwhile, the closer
the value of MG is to 1, the better the performance of the model is. After
Concentration values of gas dispersion in 11 different conditions taking the logarithm of NMSE, the closer log10 (NMSE) is to negative
were predicted. The values of input parameters in each case are shown in infinity, the smaller the error of the model is.
Table 2. It can be seen from Fig. 7 that values of NMSE obtained by CFD-
These 6 cases are mainly classified according to the wind speed. Case Spline model are the smallest, followed by CFD-Linear model, and
1–2 are the cases where the wind speeds are low, case 3–4 are the cases values of NMSE obtained by CFD-Nearest model are the largest. Further,
where the wind speeds are moderate, and case 5–6 are the cases when NMSE criterion shows a better performance of CFD-Spline model.
the wind speeds are high. FB shows the same trend. It can be seen from Fig. 8 that FB criterion
shows a better performance of CFD-Spline model. Furthermore, the
4. Results and discussion

4.1. Prediction results

It has to be reminded again that, in this paper, the simulation results


of CFD model is assumed to be the actual results of gas dispersion. The
prediction results of the fused models will be compared with the
calculation results of the classical CFD model.
Taking test case 1 as an example, the difference between the pre­
dicted concentration distribution (Cp) and CFD output (Co) was

Fig. 4. Settings of wind speed at each boundary when wind direction is


30 degrees. Fig. 5. Prediction results of Test 1 based on spline interpolation method.

5
Y. Jiang et al. Journal of Loss Prevention in the Process Industries 63 (2020) 103988

Fig. 9. MG of different models in case1-6.

Table 3
Acceptable ranges of performance criteria for air quality prediction model.
Performance criteria FB MG NMSE VG FAC2

Acceptable ranges ð 0:3; 0:3Þ ð0:7; 1:3Þ <4 < 1:6 > 0:5
Fig. 6. Enlarge part of Fig. 5.

Table 4
VG, R and FAC2 of three models in case1-6.
Case1-6 VG R FAC2

CFD-Spline 1 1 1
CFD-Linear 1 1 1
CFD-Nearest 1 1 1

performance of CFD-Linear model is better than that of CFD-Nearest


model.
Fig. 9 shows that the values of MG obtained by CFD-Spline model are
closest to 1, followed by CFD-Linear model, and the difference between
the values of MG of CFD-Nearest model and 1 is the largest. Therefore,
MG shows a better performance of CFD-Spline model as well.
In addition, the values of VG, R and FAC2 of different models in
different cases are all 1. Table 4 shows the values of VG, R and FAC2 for
case 1–6 based on different CFD-interpolation models. VG, R and FAC2
criteria show same performance of the three models.
Combining all of the above performance criteria, in case 1–6, the
Fig. 7. log10 (NMSE) of different models in case1-6. error of CFD-Spline model is the smallest, followed by CFD-Linear
model, and the error of CFD-Nearest model is the largest. Also, ac­
cording to Table 3, the performance values of the CFD-Spline model are
within the acceptable range.

4.2. Discussion

The experimental results show that the three commonly used inter­
polation algorithms can be used for the prediction of gas dispersion.
According to the performance criteria proposed by Chang and Hanna
(2004), all three CFD-interpolation models show good predictive
performance.
The difference between the VG, R and FAC2 criteria of different
models is very small. The trend of the NMSE, FB and MG criteria is the
same, which indicates that the CFD-Spline model has the best
performance.

4.3. Computational time of different gas dispersion models

The computation time of CFD model, Gaussian model and fused CFD-
interpolation models proposed in this paper is shown in Table 5. The
Fig. 8. FB of different models in case1-6. time spent on three CFD-interpolation models is similar. Among them,
the computation time of CFD-Spline model is slightly longer due to the
fitting of more complex cubic equation. CFD-Linear model and CFD-

6
Y. Jiang et al. Journal of Loss Prevention in the Process Industries 63 (2020) 103988

Nearest model have almost the same computation time. In this case, Table 5
assuming that the modeling time is not considered, it takes about 5 min Computation speed of different models.
to simulate with the CFD model and requires the support of expertise to Model Computation Time
obtain convincing results. Using CFD-interpolation model does not
CFD model 5min
require much expertise compared to using classical CFD model and does CFD-Spline model 4s
not need to solve complex Navier-Stokes equation. CFD-Spline interpo­ CFD-Nearest model 3s
lation model is about 75 times faster than CFD model. The other two CFD-Linear model 3s
interpolation models are 100 times faster than CFD model. In the same Gaussian model 0.12s

scenario, the Gaussian model takes only 0.12 s. The calculation time of
CFD-interpolation model is 25 times that of the Gaussian model. How­ Declaration of competing interest
ever, the calculation time of each CFD-interpolation model is also in
seconds, which can meet the requirements of real-time prediction in The authors declared that they have no conflicts of interest to this
emergency response. work.
Since CFD-interpolation models mainly interpolate the parameters
such as wind speed, wind direction and source strength, the calculation Acknowledgments
time of these models will not increase with the complexity of the 3D
model. For more complex computational scenarios, CFD-interpolation The authors wovuld like to express their great appreciation to their
models still can obtain results in seconds. However, the calculation teachers for their helpful suggestions on this manuscript. This work was
time of classical CFD model increases with the complexity of the 3D sponsored by Shanghai Sailing Program of China (18YF1425600) and
model and it cannot be applied to real-time prediction of emergency National Key R&D Program of China (2016YFC0801505).
scenarios.
References
5. Conclusion
Askariyeh, M.H., Kota, S.H., Vallamsundar, S., et al., 2017. AERMOD for near-road
In the emergency response of a sudden toxic gas leakage accident, it pollutant dispersion: evaluation of model performance with different emission
source representations and low wind options. Transp. Res. D Transp. Environ. 57,
is necessary to quickly and accurately predict the gas concentration
392–402. https://doi.org/10.1016/j.trd.2017.10.008.
distribution in order to define the dangerous areas and make effective Bourke, P., 1999. Interpolation methods. available at: http://paulbourke.net/miscellan
emergency decisions. CFD model is a commonly used method for eous/interpolation/.
simulating gas dispersion in complex structured urban areas. However, Chang, J.C., Hanna, S.R., 2004. Air quality model performance evaluation. Meteorol.
Atmos. Phys. 87 (1–3), 167–196. https://doi.org/10.1007/s00703-003-0070-7.
due to the complexity of CFD model, its calculation time is so long and it Dasgotra, A., Teja, G.V.V.V., Sharma, A., et al., 2018. CFD modeling of large-scale
cannot meet the requirement of real-time prediction in emergency sit­ flammable cloud dispersion using FLACS. J. Loss Prev. Process. Ind. 56, 531–536.
uations. In order to improve the calculation speed of CFD model, this https://doi.org/10.1016/j.jlp.2018.01.001.
de OB Lira, J., Padoin, N., Vilar, V.J.P., et al., 2018. Photocatalytic NOx abatement:
paper proposed a CFD-interpolation model: First, consider various mathematical modeling, CFD validation and reactor analysis. J. Hazard Mater.
possible leak situations in a certain area and use CFD model to simulate https://doi.org/10.1016/j.jhazmat.2018.07.009.
these conditions. Then the calculation results were saved as a database Dong, L., Zuo, H., Hu, L., et al., 2017. Simulation of heavy gas dispersion in a large
indoor space using CFD model. J. Loss Prev. Process. Ind. 46, 1–12. https://doi.org/
which can be quickly recalled for emergency usage. However, the 10.1016/j.jlp.2017.01.012.
database cannot include all the situations that may occur. The interpo­ Gibson, M.D., Kundu, S., Satish, M., 2013. Dispersion model evaluation of PM2. 5, NOx
lation methods can predict the future behaviors that are not encountered and SO2 from point and major line sources in Nova Scotia, Canada using AERMOD
Gaussian plume air dispersion model. Atmos. Pollut. Res. 4 (2), 157–167. https://
within the scope of the variables in the sample database. The classical
doi.org/10.5094/APR.2013.016.
CFD model has a long calculation time due to solving complex Navier- Kakosimos, K.E., Assael, M.J., 2013. Application of Detached Eddy Simulation to
Stokes equation. However, in the CFD-interpolation models proposed neighbourhood scale gases atmospheric dispersion modelling. J. Hazard Mater. 261,
653–668. https://doi.org/10.1016/j.jhazmat.2013.08.018.
in this paper, the concentrations can be obtained by simple interpolation
Kurnia, J.C., Sasmito, A.P., Mujumdar, A.S., 2014. CFD simulation of methane dispersion
calculation without solving the Navier-Stokes equation in the case and innovative methane management in underground mining faces. Appl. Math.
where the previous database has been established. Model. 38 (14), 3467–3484. https://doi.org/10.1016/j.apm.2013.11.067.
In this paper, three commonly used interpolation methods: linear Liu, X., Godbole, A., Lu, C., et al., 2015. Study of the consequences of CO2 released from
high-pressure pipelines. Atmos. Environ. 116, 51–64. https://doi.org/10.1016/j.
interpolation, cubic spline interpolation and nearest interpolation were atmosenv.2015.06.016.
proposed to combine with CFD model for real-time prediction of gas Li, X.J., Zhou, R.P., Konovessis, D., 2016. CFD analysis of natural gas dispersion in engine
leakage. Experiments have shown that CFD-Spline interpolation model room space based on multi-factor coupling. Ocean Eng. 111, 524–532. https://doi.
org/10.1016/j.oceaneng.2015.11.018.
can predict gas release with acceptable accuracy and has the highest Liu, B., Liu, X., Lu, C., et al., 2016. Computational fluid dynamics simulation of carbon
precision among the three interpolation models. What’s more, this dioxide dispersion in a complex environment. J. Loss Prev. Process. Ind. 40,
model costs less time than CFD model. It is faster than classical CFD 419–432. https://doi.org/10.1016/j.jlp.2016.01.017.
Lauret, P., Heymes, F., Aprin, L., et al., 2016. Atmospheric dispersion modeling using
model by a factor of 75. The response speed of this new model can meet Artificial Neural Network based cellular automata. Environ. Model. Softw 85, 56–69.
the time requirement of real-time prediction in emergency rescue. https://doi.org/10.1016/j.envsoft.2016.08.001.
However, there are still some problems with the model proposed in Lauret, P., Heymes, F., Forestier, S., et al., 2017. Forecasting powder dispersion in a
complex environment using Artificial Neural Networks. Process Saf. Environ. Prot.
this paper. First of all, the CFD-interpolation models proposed in this
110, 71–76. https://doi.org/10.1016/j.psep.2017.02.003.
paper only consider three main factors: wind speed, wind direction and Mack, A., Spruijt, M.P.N., 2013. Validation of OpenFoam for heavy gas dispersion
leakage rate. In the actual leakage scenario, there are various parame­ applications. J. Hazard Mater. 262, 504–516. https://doi.org/10.1016/j.
jhazmat.2013.08.065.
ters such as temperature, downwind distance, atmospheric stability,
Mishra, K.B., Wehrstedt, K.D., Krebs, H., 2014. Amuay refinery disaster: the aftermaths
etc., which will also affect the gas dispersion. Improved model consid­ and challenges ahead. Fuel Process. Technol. 119, 198–203. https://doi.org/
ering multiple factors remains to be studied. Additionally, current 10.1016/j.fuproc.2013.10.025.
research is only for a single source of leakage, and there may be multiple Ma, D., Zhang, Z., 2016. Contaminant dispersion prediction and source estimation with
integrated Gaussian-machine learning network model for point source emission in
sources of leakage in the actual situation. Future work will involve more atmosphere. J. Hazard Mater. 311, 237–245. https://doi.org/10.1016/j.
complex structured areas and more complicated leakage scenarios. jhazmat.2016.03.022.
Ma, D., Tan, W., Zhang, Z., et al., 2017. Parameter identification for continuous point
emission source based on Tikhonov regularization method coupled with particle
swarm optimization algorithm. J. Hazard Mater. 325, 239–250. https://doi.org/
10.1016/j.jhazmat.2016.11.071.

7
Y. Jiang et al. Journal of Loss Prevention in the Process Industries 63 (2020) 103988

Matacchiera, F., Manes, C., Beaven, R.P., et al., 2018. AERMOD as a Gaussian dispersion Wang, B., Chen, B., Zhao, J., 2015. The real-time estimation of hazardous gas dispersion
model for planning tracer gas dispersion tests for landfill methane emission by the integration of gas detectors, neural network and gas dispersion models.
quantification. Waste Manag. https://doi.org/10.1016/j.wasman.2018.02.007. J. Hazard Mater. 300, 433–442. https://doi.org/10.1016/j.jhazmat.2015.07.028.
Na, J., Jeon, K., Lee, W.B., 2018. Toxic gas release modeling for real-time analysis using Wang, B., Qian, F., 2018. Three dimensional gas dispersion modeling using cellular
variational autoencoder with convolutional neural networks. Chem. Eng. Sci. 181, automata and artificial neural network in urban environment. Process Saf. Environ.
68–78. https://doi.org/10.1016/j.ces.2018.02.008. Prot. 120, 286–301. https://doi.org/10.1016/j.psep.2018.09.006.
Rum, A., Landucci, G., Galletti, C., 2018. Coupling of integral methods and CFD for Zheng, X., Chen, Z., 2011. Inverse calculation approaches for source determination in
modeling complex industrial accidents. J. Loss Prev. Process. Ind. 53, 115–128. hazardous chemical releases. J. Loss Prev. Process. Ind. 24 (4), 293–301. https://doi.
https://doi.org/10.1016/j.jlp.2017.09.006. org/10.1016/j.jlp.2011.01.002.
Wang, Z., Hu, Y., Jiang, J., 2013. Numerical investigation of leaking and dispersion of
carbon dioxide indoor under ventilation condition. Energy Build. 66, 461–466.
https://doi.org/10.1016/j.enbuild.2013.06.031.

You might also like