Proactive Abnormal Emission Identification by Air-Quality-Monitoring

May 31, 2013 - Based on an available air-quality-monitoring network, to detect the possible emission sources (chemical plants) for an observed emissio...
0 downloads 0 Views 5MB Size
Article pubs.acs.org/IECR

Proactive Abnormal Emission Identification by Air-Quality-Monitoring Network Tianxing Cai,† Sujing Wang,‡,§ Qiang Xu,*,† and Thomas C. Ho† †

Dan F. Smith Department of Chemical Engineering and ‡Department of Computer Science, Lamar University, Beaumont, Texas 77710, United States § Department of Computer Science, University of Houston, Houston, Texas 77004, United States ABSTRACT: Chemical facilities, where large amounts of chemicals and fuels are processed, manufactured, and housed, are at high risk to originate air emission events, including intensive flaring and toxic gas release caused by various uncertainties such as equipment failure, false operation, natural disaster, or terrorist attack. Based on an available air-quality-monitoring network, to detect the possible emission sources (chemical plants) for an observed emission event, so as to support diagnostic and prognostic decisions in a timely and effective manner, a systematic method for abnormal emission identifications should be employed. In this article, a systematic methodology for such applications has been developed. It includes two stages of modeling and optimization work: (1) the determination of background normal emission rates from multiple emission sources and (2) multiobjective optimization for emission-source identification and quantification. This method not only can determine emission source location, starting time, and time duration responsible for an observed emission event, but can also estimate in reverse the dynamic emission rate and total emission amount from an accidental emission source. It provides valuable information for investigations of accidents and root-cause analysis for emission events; meanwhile, it helps evaluate the regional air-quality impact caused by such emission events as well. Case studies including the detection of a real SO2 emission event are employed to demonstrate the efficacy of the developed methodology. plant in Louisville, KY, fired calcium carbide and produced large amounts of inhalation hazardous gases.5 The air-quality impacts from chemical plant emission events can be serious to both local communities and their surrounding environments. One of the major concerns is the exposure of acute or short-term toxicity. Release of acutely toxic contaminants, such as SO2 and chlorine, would likely be transported to a populated area and pose an immediate threat to the public health and environment quality. Generally, plant personnel should document and report emission details in response to an emission event, so that valuable information on hazardous release rate, possible transportation speed and direction, and potential harmful impacts on exposed populations and ecosystems can be estimated to support responsible decision making. Because such responsible decisions are highly critical, independent supporting information such as real-time measurements from a local air-quality-monitoring network is vital , especially in industrial zones that are populated heavily by various chemical facilities. A local air-quality-monitoring network can measure and record multiple pollutant concentrations simultaneously and raise an alarm for dangerous events in real-time fashion. Meanwhile, based on measurement data from each monitoring station plus regional meteorological conditions during the event time period, a monitoring network can help estimate possible emission source locations or even their emission rates. This inverse characterization of abnormal emission sources is very valuable

1. INTRODUCTION Chemical facilities, where large amounts of chemicals and fuels are processed, manufactured, and housed, are at high risk to give rise to air emission events. Normal plant emissions are those routine emissions that are expected during plant normal operations, and they are regulated and included in planned emissions profiles. Abnormal plant emissions are significant amounts of emissions during periods of plant startup, shutdown, or malfunction. According to the U.S. Environmental Protection Agency’s (EPA) 40 Code of Federal Regulation, malfunctions refer to any “sudden, infrequent, and not reasonably preventable failure of air pollution control equipment, process equipment, or a process to operate in a normal or usual manner”.1,2 Air emission events can also be caused by severe process upsets due to planned operations such as plant scheduled turnarounds (startup or shutdown). For example, an olefin plant with ethylene productivity of 544000 ton/year can easily flare about 2268 tons of ethylene during a single startup,3 resulting in at least 18 tons of CO, 3.4 tons of NOx, and 45.4 tons of highly reactive volatile organic compounds (HRVOCs, defined in Texas air-quality regulation as ethylene, propylene, isomers of butene, and 1,3butadiene). If all other flaring species (ethane, propylene, propane, etc.) are also taken into account, tremendous air emissions can be produced within a short-time period. Chemical plant emission events can also be caused by uncontrollable and unpredictable uncertainties, such as emergency shutdowns, natural disasters, or terrorist attacks. For example, an oil refinery in eastern Japan exploded with huge amounts of toxic emissions as a result of the earthquake and tsunami that occurred on March 11, 2011.4 In another emission event on March 22, 2011, a blast at a carbide © XXXX American Chemical Society

Received: February 21, 2013 Revised: May 29, 2013 Accepted: May 31, 2013

A

dx.doi.org/10.1021/ie400568c | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX

Industrial & Engineering Chemistry Research

Article

Table 1. Classification of Abnormal Emission Patterns

in an average long-time period based on measurements from multiple monitoring stations. This means that their emissions are assumed to be under steady-state conditions and their values are treated as constants. Therefore, studies on reverse modeling for abnormal emission identifications with the consideration of dynamic emission rates of point emission sources are still lacking. Note that abnormal emissions are quite different from normal emissions and are most often associated with emergency situations because of their large emission amounts and fast response requirements. The timely detection of the emission source(s), emission rate, and starting time will provide a new perspective to better understand an emission event and also help quickly

to all stake holders, including government environmental agencies, chemical plants, and residential communities. In the bulk of previous research, inverse modeling ideas were originated by the use of atmospheric dispersion models, which have normally been used in the forward modeling problem to determine downwind contamination concentrations with given meteorological conditions and emission rates. The “Gaussian plume model” is an approximate analytical method for the calculation of air pollutant concentrations in the downwind area from point-source emissions.6−18 Even though inverse modeling methods based on the Gaussian plume model have been reported,19−21 they are generally used to estimate emission rates of point sources B

dx.doi.org/10.1021/ie400568c | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX

Industrial & Engineering Chemistry Research

Article

predict its air-quality impact. Also note that, although emission reports from the source of the event will eventually become available, they might not be obtained in time and the emission rate will most likely be an approximated constant value instead of a dynamic profile. In practice, dynamic emission profiles can be very useful. For example, they can help schedule the activities of multiple chemical plants to minimize regional airquality impacts.22 Therefore, the study of the identification of abnormal emissions has great significance to support quick and smart decision making for multiple stakeholders for the sake of regional air quality. In this article, a systematic methodology for abnormal emission identifications has been developed. It includes two stages of modeling and optimization work: (1) the determination of background normal emission rates from multiple emission sources and (2) multiobjective optimization for emission source identification and quantification. This approach not only can determine emission source location, starting time, and time duration responsible for an observed emission event, but can also estimate, in reverse, the dynamic emission rate and the total emission amount from the accidental emission source. It provides valuable information for accident investigations and root-cause analysis for emission events and also helps evaluate regional air-quality impacts caused by such an emission event.

Table 2. Key Parameter Characterization for Various Emission Patterns



⎤ ⎥ ⎥ 1 + exp( d | t − e | + d | t − f | ) i,k i,k i,k i,k ⎦ k∈ Ki ⎣

∑ ⎢⎢

bi,k

single acute peak single plain peak single inclining peak multiple acute peaks multiple plain peaks multiple inclining peaks

0 0 1 0 0 1

ei,k and f i,k ei,k ei,k ei,k ei,k ei,k ei,k

= f i,k ≠ f i,k ≠ f i,k = f i,k ≠ f i,k ≠ f i,k

dimension of set Ki 1 1 1 number of peaks number of peaks number of peaks

according to the abnormal air-quality measurements from an available monitoring network, so as to support diagnostic and prognostic decisions in a timely and effective manner. The outcome of the developed methodology should provide information on the emission source location, starting time, time duration, total emission amount, and dynamic emission rate and pattern from the abnormal emission sources. The methodology first determines the normal background emission rates for a given list of candidate emission sources in the region. Next, an optimization model is employed for reverse emission source detection based on abnormal air-quality measurements. For clarity, the problem statements are summarized next. The problem relies on the following assumptions: (1) An air-quality event in a region is caused by abnormal emissions from one and only one emission source based on a given list of candidate emission sources, whose abnormal emission pattern is one of those shown in Table 1. (2) Each candidate emission source has a constant emission rate under normal operating conditions. (3) Emission transportation follows a Gaussian dispersion model, and there is no secondary consumption or generation of the pollutant during its air transportation. (4) When the emitted pollutant reaches the ground through dispersion, it is absorbed, that is , there is no pollutant reflection from the ground during its air transportation. (5) Meteorological conditions (e.g., local wind direction and wind speed) during the considered scheduling time horizon are constant (or nearly constant) in the region. The following information is given: (1) spatial locations of each emission sources and monitoring stations; (2) emission source stack parameters, such as stack height and outlet temperature; (3) dynamic monitoring results at each monitoring station; and (4) meteorological conditions in the studied region during the event time period. Finally, the information to be determined includes (1) the identity of the emission source that caused the investigated emission event, (2) the starting and ending times associated with the emission event, and (3) the emission pattern and dynamic emission rate for the identified emission source.

2. MODELING OF ABNORMAL EMISSION PATTERNS Chemical plant emissions include normal routine emissions and abnormal emissions. Abnormal emissions are caused by planned, unpredictable, or uncontrollable events that significantly upset or damage plant equipment, process facility, ventilation system, or flare control system, which, in turn, generates large amounts of emissions. Emissions due to process turnaround operations (e.g., startup and shutdown) as well as equipment leakage, rupture, or explosions are typical examples. Conceivably, abnormal emissions from chemical plants exhibit several typical dynamic emission profiles. They can be approximately classified into seven typical emission patterns, including single acute peak, single plain peak, single declining peak, multiple acute peaks, multiple plain peaks, multiple declining peaks, and multiple hybrid peaks. Table 1 summarizes the general shapes and characterizations of each abnormal emission pattern and also provides possible root causes related to each of these abnormal emission patterns. To quantitatively characterize each abnormal emission patterns in Table 1, a general mathematical model was developed miD(t ) =

dynamic emission pattern

ai , k (bi , k t + ci , k)

(1)

miD(t)

where represnts the dynamic emission rate from emission source i at time t; k is the emission peak index of mDi (t); Ki represents the set of emission peak index of mDi (t); and ai,k, bi,k, ci,k, di,k, ei,k, and f i,k are corresponding parameters. Coupling these parameters together can help characterize three major elements of a dynamic emission pattern: number of emission peaks, peak duration, and peak decreasing trend. As examples, Table 2 provides the value ranges for the key parameters that could be used to generate corresponding dynamic emission patterns.

4. GENERAL METHODOLOGY In this section, the methodology framework is first introduced. Then, two key technologies associated with this methodology, namely, normal emission rate determination and abnormal emission source identification, are presented.

3. PROBLEM STATEMENT Based on the aforementioned model, the studied problem was to develop a systematic methodology to detect, in reverse, the emission source from a list of candidates (local chemical plants) C

dx.doi.org/10.1021/ie400568c | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX

Industrial & Engineering Chemistry Research

Article

Figure 1. Methodology framework.

4.1. Methodology Framework. The methodology includes two stages of work. As illustrated in Figure 1, input parameters in the first modeling stage include geographical information (locations of all possible emission sources and monitoring stations), meteorological conditions (e.g., wind direction, wind speed, atmospheric stability), measurements at each monitoring station, and emission source data (e.g., stack height, exit diameter, and outlet temperature). The next step is to map locations of candidate emission sources and monitoring stations onto a rectangular coordinate system (see the Appendix). Then, the first-stage modeling aims at the determination of normal emission rates from every emission source. The task is accomplished through a regression model based on the Gaussian dispersion model to minimize the sum of squared error (SSE) between the model-calculated results and the monitoring results from multiple monitoring stations. Because the normal emission rate of each emission source is the background emission during plant steadystate operating conditions, it is also called the steady-state emission rate. In the second stage, when an emission event has been observed through measurements from regional monitoring stations,

the Gaussian-dispersion-model-based optimization model for emission pattern identification is set up and solved. The potential emission source for this emission event and its dynamic emission rate are identified, which generates a multiobjective optimization problem to minimize (1) the sum of squared relative error (SSRE) between the model-calculated and monitoring results during the time period of the emission event and (2) the sum of relative ratio (SRR) of the calculated abnormal emission rate to the normal emission rate for all emission sources. The minimization of the first objective is obvious. The reason for setting up the second objective is to avoid multiple solution possibilities that one emission event might be caused by different emission rate scenarios from different emission sources. Thus, adding the objective of minimizing SRR is approximately equivalent to identifying a solution that results in the observed emission event with the minimum total amount of emissions from an emission source. Finally, the obtained optimization results need to be validated by air-quality simulation with the Gaussian dispersion model. 4.2. Determination of the Normal Emission Rate. This regression model is to identify the normal emission rate for each candidate emission source by minimizing SSE between the D

dx.doi.org/10.1021/ie400568c | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX

Industrial & Engineering Chemistry Research

Article

an air-quality issue and its dynamic emission rate. It simultaneously minimizes two objective functions (ϕ2 and ϕ3). 4.3.1. Objective Functions. One of these objective functions is given by

model-predicted results and the monitoring results from multiple monitoring stations under normal emission status (without emission events). 4.2.1. Objective Function. The objective function is given by ϕ1 = min S mi

∑ ∑ (Cj̅ − C jS,t )2 t∈T

D

ϕ2 =

(2)

j∈J

where j represents the index of monitoring stations grouped by set J and TS represents a selected steady-state time set when each emission source has a normal emission rate. TS contains multiple time instants indexed by t. C̅ j and CSj,t represent the modelcalculated and measured pollutant concentrations, respectively, at the jth monitoring station at time t. Equation 2 suggests that the objective function is to minimize the SSE between C̅ j and CSj,t. The model-calculated pollutant concentration at the jth monitoring station (C̅ j) should be the sum of C̅ i,j values from all of the emission sources, which can be formulated as Cj̅ =

∑ Ci̅ ,j ,

where Cj,t represents the model-calculated result and is the monitoring result at station j at time instant t within dynamic state set TD (with emission event), {a, b, c, d, e, f}i,k is the set of associated parameters of dynamic emission pattern of source i (see eq 1), which are manipulated by the optimization model; and yi is the binary variable to determine whether source i has the abnormal emission event or not, such that yi = 1 if source i is the abnormal emission source and yi = 0 otherwise. This objective function is to minimize the SSRE between the modelcalculated results and monitoring results from multiple monitoring stations in dynamic situations characterized by the dynamic state set TD, which consists of a finite arithmetic progression distributed within the event time window [TL, TU] as shown in

(3)

4.2.2. Dispersion Transportation Principle. Note that C̅ i,j should be calculated as Ci̅ , j =

miSf (Zi , j , Hi) 2UiπσYi ,jσZi ,j

⎡ ⎛ ⎞2 ⎤ ⎢ 1 ⎜ Yi , j ⎟ ⎥ exp⎢ − ⎜ ⎟ ⎥, ⎢⎣ 2 ⎝ σYi ,j ⎠ ⎥⎦

⎧ ⎫ TU − T L T D = ⎨T L + nδi n ≤ , n ∈ N *⎬, δi ⎩ ⎭

∀ i ∈ I, j ∈ J









∀i∈I (6)

(4)

where n belongs to the positive integer set (N*) and δi is a given minimum time step for the ith emission source. The second objective of the optimization model is to minimize the SRR of the emission rate, which is the sum of the relative ratio of mDi (t) to mSi as described given by

where C̅ i,j is the pollutant concentration at the jth monitoring station caused by the emission from the ith emission source; Yi,j is the projection of di,j along the Y direction (here, the X direction is the same as the wind direction, and the Y direction is horizontally perpendicular to X direction); Zi,j is the groundheight difference between the ith emission source to the jth monitoring station; Hi is the plume height above the ground for source i; σYi,j and σZi,j are the standard deviations of the emission plume’s probability distribution function along the Y and Z directions, respectively; f(Zi,j,Hi) is a function of Zi,j and Hi; and mSi represents the constant emission rate at the ith emission source under normal conditions. Equation 4 represents pollutant dispersion from emission sources to monitoring stations under the impact of meteorological conditions. The details associated with eq 4 can be found in the Appendix. The related equations and procedures for calculating the parameters in eq 4 are included in the Gaussian dispersion model6−16 and are introduced in the Appendix. 4.3. Determination of Abnormal Emission Source. The target of this task is to identify the potential emission source for

Ci , j , t

(5)

CDj,t

∀j∈J

i∈I

min

{a , b , c , d , e , f }i , k yi

⎛ C − C D ⎞2 j,t j,t ⎟ ∑ ∑ ⎜⎜ ⎟ D C D ⎠ j,t t∈T j∈J ⎝

ϕ3 =

min

∑∑

{a , b , c , d , e , f }i , k t∈T D i∈I yi

miD(t ) miS

(7)

It helps identify the solution with the minimum deviation from the normal emission rates of all of the emission sources during the dynamic time span. 4.3.2. Model Constraints. The real concentration at the jth monitoring station at time t (Cj,t) during the emission event should be the cumulative result of Ci,j,t from all of the emission sources, which is expressed by the equation Cj , t =

∑ Ci ,j ,t ,

∀ j ∈ J , ∀ t ∈ TD (8)

i∈I

Note that Ci,j,t should be calculated by the equation

⎧C̅ , t < TSi + τi , j ⎪ ⎪ ⎡ ⎛ ⎞2 ⎤ = ⎨ [(1 − y )miS + ymiD(t − TSi − τi , j)]f (Zi , j , Hi) 1 ⎜ Yi , j ⎟ ⎥ ⎢ i i ⎪ exp⎢ − ⎜ ,t ≥ TSi + τi , j 2UiπσYi ,jσZi ,j 2 ⎝ σYi ,j ⎟⎠ ⎥⎥ ⎪ ⎢ ⎣ ⎦ ⎩

where TSi is the starting time of the e mission event from the ith emission source and τi,j is the emission transportation delay from the ith emission source to the jth monitoring station. mDi (t − TSi − τi,j) is the time delay format of the dynamic emission rate m iD (t). Again, the associated parameter calculations of eq 9 can be found in the Appendix.

∀ i ∈ I , j ∈ J , t ∈ TD (9)

Note that, in eq 9, for t < TSi + τi,j, Ci,j,t is equal to C̅ i,j because the potential abnormal emissions have not impacted the jth monitoring station yet because of the time delay. Thus, Ci,j,t supposedly keeps the normal emission impact of C̅ i,j, which is identified in the first stage of modeling work. For t ≥ TSi + τ i,j, Ci,j,t can be calculated by the standard E

dx.doi.org/10.1021/ie400568c | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX

Industrial & Engineering Chemistry Research

Article

⎧ ⎫ ciU, k − ciL, k ⎪ ⎪ L ⎬, + δ ≤ ∈ * ci , k ∈ ⎨ c n c n , n N i , k i , k ⎪ ⎪ δci , k ⎩ ⎭ ∀ i ∈ I , k ∈ Ki

di , k

(14)

⎫ ⎧ diU, k − diL, k ⎪ ⎪ L ⎨ , n ∈ N *⎬ , ∈ ⎪di , k + nδdi , k n ≤ ⎪ d δ i,k ⎭ ⎩

∀ i ∈ I , k ∈ Ki

(15)

Table 3. Plume and Stack Parameters for Each Emission Source in Case Study 1 chemical emission sources stack stack stack stack

Figure 2. Spatial scope of case study 1.

height Hi (m) exit temperature Ts,i (K) exit velocity Vs,i (m/s) exit diameter Ds,i (m)

E1

E2

E3

E4

E5

80 480 17.5 1.6

110 400 13.0 1.9

95 460 15.6 1.5

100 440 14.2 1.7

105 430 16.1 1.8

Gaussian dispersion model with a manipulated emission rate associated with yi: If the ith emission source does not involve an abnormal emission event (yi = 0), the emission rate of source i retains the constant normal emission rate of mSi ; otherwise (yi = 1), the emission rate of source i is characterized by the dynamic emission pattern mDi (t − TSi − τ i,j) quantified by eq 1. The pollutant transportation time (τi,j) from the ith source to the jth monitoring station is calculated as di⃗ , j·U⃗

, ∀ i ∈ I, j ∈ J 2 (10) || U⃗ || It is actually the projection of the spatial vector di,j⃗ (indicating the distance and direction from emission source i toward monitoring station j) to the wind vector of U⃗ (indicating the wind direction and speed), namely, (d⃗i,j·U⃗ )/(||U⃗ ||), then divided by the magnitude of the wind speed ||U⃗ ||. Also note that the manipulated variable TSi equals the underdetermined coefficient ei,1 in eq 1 τi , j =

TSi = ei,1 ,

∀i∈I

Figure 3. Pollutant concentration measurements from local monitoring stations.

(11)

Equation 12 suggests one and only one potential source has caused the emission event

∑ yi = 1 i∈I

(12)

The optimization to inversely identify the dynamic emission rates of abnormal emission sources is a highly nonlinear MINLP (mixed-integer nonlinear programming) problem, especially because of eqs 1 and 9. To reduce the problem complexity, the manipulated variables ai,k, ci,k, di,k, ei,k, and f i,k were individually discretized into finite arithmetic progressions distributed within their upper and lower bounds, as follows ai , k

⎫ ⎧ aiU, k − aiL, k ⎪ ⎪ L ⎨ ∈ ⎪ai , k + nδai , k n ≤ , n ∈ N *⎬ , ⎪ δai , k ⎭ ⎩

∀ i ∈ I , k ∈ Ki

(13)

Figure 4. Identified dynamic emission profile of emission source E3. F

dx.doi.org/10.1021/ie400568c | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX

Industrial & Engineering Chemistry Research

Article

Figure 5. Comparison between simulation and measurement results at (a) S1, (b) S2, (c) S3, and (d) S4 for case study 1.

⎫ ⎧ eiU, k − eiL, k ⎪ ⎪ L ⎬, ei , k ∈ ⎨ e n e n , n N + δ ≤ ∈ * i , k i , k ⎪ ⎪ δei , k ⎭ ⎩ ∀ i ∈ I , k ∈ Ki

fi , k

multiobjective genetic algorithm (MGA) provided by the Global Optimization Toolbox of MATLAB (MathWorks, R2010a). The major reasons for selection of this solver include the following: (1) MGA can directly obtain Pareto frontier of the raised two objectives from all possible dynamic emission patterns under different optimal scenarios. (2) Because the airquality impact of the entire local area might be investigated, the whole region needs to be partitioned into a large number of grid cells, and the emission concentration of each cell must be calculated. Thus, the optimization involves many simulations, causing the simulation duty during the optimization to be very heavy. (3) The application of the Global Optimization Toolbox in MATLAB greatly facilitates the seamless connection between optimization and simulation for data transfer and result presentation (e.g., three-dimensional visualization during the solution procedure). (4) The decision variables involve discretized forms of ai,k, ci,k, di,k ei,k, and f i,k, which are also easy to handle by MGA. It should be noted that the MGA solver used in this article cannot guarantee the solution with global optimality.

(16)

⎧ ⎫ fiU, k − fiL, k ⎪L ⎪ , n ∈ N ⎬, ∈ ⎨fi , k + nδfi , k n ≤ ⎪ ⎪ δfi , k ⎩ ⎭

∀ i ∈ I , k ∈ Ki

(17)

where n belongs to the positive integer set (N*) and δai,k, δci,k, δdi,k, δei,k, and δf i,k are given minimal changes of associated parameters for the ith emission source. 4.3.3. Solution Algorithm. The optimization to reversely determine the dynamic emission rate from an abnormal emission source is a high nonlinear multiobjective MINLP problem, which is difficult to solve with current deterministic MINLP solvers. Thus, the solution algorithm used in this study is a G

dx.doi.org/10.1021/ie400568c | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX

Industrial & Engineering Chemistry Research

Article

Figure 6. Simulated pollutant distributions based on the identified abnormal emissions from E3.

distributed in a square region (30 km × 30 km). The entire region is gridded, and the edge length of each grid cell is 1 km. The surface wind blows from the southwest to the northeast as shown in Figure 2. A common pollutant emitted from these five emission sources is monitored by the monitoring stations through hourly measurements. The plume and stack parameters of the five emission sources are given in Table 3. In this case study, the step changes of δai,k,

5. CASE STUDIES To demonstrate the efficacy of the developed systematic methodology, two case studies including the detection of a real SO2 emission event were examined. 5.1. Case Study 1. As shown in Figure 2, case study 1 involves five chemical emission sources (E1−E5, represented by red dots) and four monitoring stations (S1−S4, represented by green dots) H

dx.doi.org/10.1021/ie400568c | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX

Industrial & Engineering Chemistry Research

Article

Figure 7. Simulated pollutant contours based on the identified abnormal emissions from E3.

δci,k, δdi,k, δei,k, and δf i,k are set as 0.01; the temporal computational resolution δi is 0.1 h. The monitoring results show that a high-concentration event occurred on an overcast evening

with mild wind. As shown in Figure 3, the monitoring results at station S3 reached a concentration of 98.1 ppb at time 01:00 and 91.9 ppb at time 02:00. Even though the concentration decreased I

dx.doi.org/10.1021/ie400568c | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX

Industrial & Engineering Chemistry Research

Article

801.1 kg/h at time 01:00 and 747.8 kg/h at time 02:00. After that, there was a deep reduction of the emission rate: It decreased to 220.1 kg/h at 03:00 and returned again to 581.2 kg/h at time 04:00. Finally, after it decreased to 90.1 kg/h, and the emission rate was restored to the normal emission value. Therefore, the dynamic emission pattern of this emission event had multiple acute peaks. Based on the calculated dynamic emission rate from source E3, the Gaussian dispersion model was used again to validate the optimization results and examine the air-quality impact on the studied region. First, the calculated dynamic concentrations at the monitoring stations (S1 and S2) were compared with the monitor measurements, which are shown in Figure 5. This comparison demonstrates that the optimization−solution-based simulation results and the historical measurements matched very well. For alarm monitoring station S3, there were only small overestimations for the concentration at 03:00 and 05:00 and a very slight underestimation at time 04:00 with a maximum error of 2 ppb. For the nonalarm stations (S1, S2, and S4), the maximum errors were within 1 ppb. Second, the air-quality impacts to the entire region due to this abnormal emission from source E3 were also analyzed. Figure 6 shows the spatial and temporal distributions of the pollutant concentration. The dynamic emission event started from 23:00, and the first peak concentration reached 98.1 ppb, which was maintained for 2 h until it slowly decreased to 91.9 ppb. Then, it declined sharply to 24.6 ppb. The second peak profile appeared at time 04:00 with a concentration of 72.1 ppb. The time duration of the second peak was much shorter than that of the first, as it lasted for only 1 h. The peak concentration decreased to 9.2 ppb and then gradually reached the background value. Thus, the entire emission event was sustained for 7 h. Figure 7 shows a dynamic contour plot of pollutant concentration during the emission event, revealing the dispersion scope of the emission event, which involved a large area along the wind direction. Both Figures 6 and 7 provide valuable air-quality impact information for this emission event. Also, thanks to the developed methodology, the investigation of future root causes of the emission event in plant E3 as well as the potential impacts on regional air-quality in the near future becomes viable.

to 24.6 ppb at time 03:00, it increased again to 72.1 ppb at time 04:00. At time 05:00, the concentration was 9.2 ppb, which is still higher than the normal background measurements. Thus, abnormal emissions from one of the five emission sources had to be detected. During the event time period, the lapse rate was 4 K/km, the wind speed was 1.6 m/s at 10-m height, and the ambient temperature was 20 °C. For the investigation, the entire time period was separated into two parts: the steady-state time period (19:00−23:00) and the dynamic time period (00:00−05:00). The data in the steadystate period were used to determine the normal emission rates for each emission source as described in section 4.2. The results for mS1−mS5 are 10.8, 10.5, 10.1, 9.2, and 9.8 kg/h, respectively. The next step is to identify the dynamic emission rate of the abnormal emission source. The results of model optimization by section 4.3 indicated that E3 is the source that caused the emission event, and its dynamic emission pattern was also identified, as shown in Figure 4. This figure shows that the emission rate increased to 88.4 kg/h from the normal emission rate at time 00:00. As time continued, the emission rate reached

Figure 8. Spatial scope of case study 2.

Figure 9. SO2 concentrations measured at (a) S1 and (b) S2. J

dx.doi.org/10.1021/ie400568c | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX

Industrial & Engineering Chemistry Research

Article

Figure 10. Identified dynamic emission profile of emission source E1.

Figure 11. Comparison between simulated and measured SO2 concentrations at S1 and S2 for case study 2.

5.2. Case Study 2. The second case study is about the detection of a real industrial SO2 emission event. The hourly SO2 monitoring results and the meteorological conditions (such as temperature, wind speed, and wind direction) come from the web site of a state environment agency and U.S. EPA, whereas spatial locations of emission sources and motoring stations were acquired from Google Earth. For reasons of confidentiality, the detailed emission source and monitoring information in terms of names and locations are ignored in this article. The geological distribution of emission sources and monitoring stations is shown in Figure 8. The studied SO2 emission event was detected at monitoring station S1 with the highest SO2 concentration of 125.9 ppb at time 13:00 (see Figure 9a). The corresponding SO2 concentration measurements at monitoring

station S2 were shown in Figure 9b. The measurement data series was classified into two parts: the data between 00:00 and 06:00 were considered as steady-state measurements, whereas the data between 07:00 and 20:00 were considered to be dynamic. The meteorological data collected during the event include an ambient temperature of 272.7 K, a wind direction of 354°, and a wind speed of about 8.4 m/s at 10-m height. Steady-state SO2 measurements were used to determine the normal emission rates of three emission sources according to the first-stage regression model of the developed methodology. The results indicated that the normal emission rates of sources E1−E3 were 2.73, 0.38, and 1.25 kg/h, respectively. The second-stage modeling was then performed, from which the emissions from E1 were identified as the root cause of the SO2 K

dx.doi.org/10.1021/ie400568c | Ind. Eng. Chem. Res. XXXX, XXX, XXX−XXX

Industrial & Engineering Chemistry Research

Article

Table 5. Power-Law Exponents (p) for Six Atmospheric Stability Categories location

location

stability

urban area

rural area

stability

urban area

rural area

A B C

0.15 0.15 0.20

0.07 0.07 0.10

D E F

0.25 0.30 0.30

0.15 0.35 0.55

Table 6. Meteorological Condition Defining Pasquill Stability Categories daytime incoming solar radiation U10 surface wind speed at 10-m height 6 m/s

Figure 12. Illustration of θW, θA, Xi,j, and Yi,j.22

Table 4. Empirical Expressions for Dispersion Coefficients for Urban Sites stability A, B C D E, F

σYi,j (m) 0.32Xi,j(1 0.22Xi,j(1 0.16Xi,j(1 0.11Xi,j(1

+ + + +

0.0004Xi,j)−0.5 0.0004Xi,j)−0.5 0.0004Xi,j)−0.5 0.0004Xi,j)−0.5

strong moderate A A−B B C C

A−B B B−C C−D D

slight B C C D D

nighttime cloud cover >1/2