Anda di halaman 1dari 18

GCPS 2013 __________________________________________________________________________

A Hybrid Model for Characterizing the Source of Hazardous Material Release


Shahryar Khajehnajafi, PhD Anjana Meel, PhD SAFER Systems LLC 5284 Adolfo Road, Suite 100 Camarillo, CA 93012 shah@safersystem.com Copyright SAFER Systems LLC, 2013

Prepared for Presentation at American Institute of Chemical Engineers 2013 Spring Meeting 9th Global Congress on Process Safety San Antonio, Texas April 28 May 1, 2013 UNPUBLISHED

AIChE shall not be responsible for statements or opinions contained in papers or printed in its publications

GCPS 2013 __________________________________________________________________________

A Hybrid Model for Characterizing the Source of Hazardous Material Release


Shahryar Khajehnajafi, PhD Anjana Meel, PhD SAFER Systems LLC 5284 Adolfo Road, Suite 100 Camarillo, CA 93012 shah@safersystem.com Keywords: hazardous chemicals, chemical spill, source of release, source of hazardous material, plume model, Source Area Locator (SAL), source characterization, Gaussian plume model, Bayesian method, Bayesian inference, hybrid model, odor source, source location estimation, burro

Abstract
On Sept 10, 2012, Los Angeles Fire department and 911 operators received many complaints about a strong foul smell. The complaint covered an area of 10,000 square miles. There were many speculations for the source of the release; sewage, chemical spill, pipeline carrying natural gas. However, after many hours of investigation by South Coast Air Quality Management it was discovered that the source of the foul odor was dead fish at the Salton Sea churned up by a thunderstorm that had occurred that evening. The Salton Sea is a 376-square-mile saltwater lake about 150 miles southwest of Los Angeles. The odor remained strong for up to 150 miles downwind of the source, and quite unusual. Fortunately the above case did not pose any health effects. It is common for industrial sites carrying hazardous chemicals to receive calls from surrounding communities complaining of a chemical presence in the ambient air. Following a complaint, industrial sites must work promptly check and verify if the odor source is related to a chemical release on their property, especially when there are other industrial sites in the vicinity, for mitigating and reducing the extent of the possible hazard. Locating the source and amount of release for predicting the geographical extent of the exposure in a timely manner is, therefore, essential to take an effective response action for evacuation or shelter-in-place. With only basic knowledge of meteorological information; wind speed, wind direction, and stability along with two or more gas concentration measurements, it is possible to quickly frame the zone of chemical release. This is an inverse problem to the usual forward dispersion calculation in which the source location and release rate is known and the concentrations at sensors positions are predicted. There are many publications in the area of source characterization; location, and amount of chemical being released. The two most cited methods are: Bayesian inference using Markov Chain Monte Carlo (MCMC) sampling methodology, and adjoint advection- diffusion equation. The Bayesian methodology fits well for application to
Note: Do not add page numbers. Do not refer to page numbers when referencing different portions of the paper

GCPS 2013 __________________________________________________________________________

atmospheric dispersion where ambient and gas sensor data are stochastic in nature, however, it takes a long time to converge. In this paper, we briefly describe a simple least-square based method for source characterization using Gaussian plume model. We apply the Bayesian method to address the same problem and compare the results of the two methodologies for few selected release scenarios as well as a set of field data. Subsequently, we will present a hybrid model using Bayesian inference which utilizes the result of the simple least- square model as a seed for narrowing the search domain and reducing the time for locating source of chemical release. We present the improvement in the run time using the hybrid model which makes it suitable for emergency response and provides a reasonably accurate prediction of source characterization.

1 Introduction
In this paper, the focus is on estimating the source area location of a hazardous release using the sensor measurements obtained from sensors located throughout a facility. If parameters of a system are known, the outcomes of measurements can be predicted. That is known as the forward problem. On the contrary, estimation of the values of the parameters of the system using the results of the measurements is called a backward problem. Identifying source parameters for location, quantity of release using sensor measurements is a backward problem. Tarantola (2005) discussed the inverse problem theory and methods for model parameter estimation in great detail. Usually a forward problem has a unique solution, but the inverse problem can have multiple solutions. Therefore, in an inverse problem, any apriori information should be utilized explicitly and uncertainties in the data should be handled carefully. When nonlinearities are available, it is important to describe uncertainty and apriori information using probability distribution. Bayesian theory based methods can be used to describe uncertainty. When systems are not nonlinear, least-square methods are considered very efficient. Both leastsquare methods and Bayesian Theory based methods have their own advantages and disadvantages. Least-square method usually minimizes the error between measurements and prediction to calculate the system unknowns while Bayesian theory based methods updates the prior probability distribution of unknown parameters using the sensor measurement data to obtain a posterior distribution of the unknown parameters, which provides the updated information of the system. Thus the source area location and release rate is indeed a very challenging and nonlinear problem, even for identifying the source area from a single sensor concentration measurement. Even for a single sensor measurement, there can be multiple solutions for release location all of which may lie on a line/curve with varying distance and release rates. For example, a large release rate at a further distance from a sensor may result in the same concentration values observed by a fixed sensor as a small release rate at a location closer to the sensor. When more than one sensor is involved, the problem becomes even more complicated. The Bayesian theory has been used by several researchers for estimating the source area location. Yee et al. (2007, 2008) developed a Bayesian probabilistic inferential solution for source reconstruction (both location and emission rate) from a limited noise concentrated data and the methods were tested for JU2003 field study in Oklahoma City (urban dispersion) and the ETEX

GCPS 2013 __________________________________________________________________________

(long-range dispersion on continental scales). Kim et al. (2007) developed a source reconstruction algorithm using dynamic Bayesian inference and applied it on Yonggwang atmospheric tracer experiment in Korea to show that the posterior modes of the release point and the released source rate for this experiment converged to their true values. Keats (2009) developed Bayesian inference technique for source determination in atmospheric environment. Adjoint (backward) dispersion models were used to reduce the computational effort required from calculating one [forward] plume per possible source configuration to calculating one [backward] plume per detector. Markov chain Monte Carlo (MCMC) is used to efficiently sample from the posterior distribution for the source parameters; both the Metropolis- Hastings and hybrid Hamiltonian algorithms are used. These algorithms were applied on a full-scale field experiment (Joint Urban 2003) in Oklahoma City and the results from the algorithms were also compared with the US Environmental Protection Agencys standard CMB model using a test case involving PM data from Fresno, California. Neuman et al. (2005) used an urban puff model [the Defense Science Technology Laboratory's (Dstl) Urban Dispersion Model (UDM)] for prediction plume evolution in complex urban geometries that enabled rapid urban dispersion simulations by combining traditional Gaussian puff modeling with empirically deduced mixing and entrainment approximations. In their work, a preliminary reconstruction of an atmospheric release event was demonstrated using stochastic sampling algorithms and Bayesian inference together with the rapid UDM urban puff model based on point measurements of concentration. This model was applied to a prototype isolated building and for observations and flow conditions taken during the Joint URBAN 2003 field campaign at Oklahoma City. Ideally, the urban puff model is used to approximate the source location and strength. The more accurate CFD model can then be used to refine the solution. Zheng et al. (2011) reviewed inverse calculation approaches for source determination and classified them into probability modeling methods and optimization modeling methods. According to them, probability modeling methods are demonstrated to be insufficient during emergency responses due to their lacking of enough prior information of unknown parameters, while optimization modeling methods are efficient and become a new trend in source determination. Sharan et al. (2011) applied least square data assimilation for identification of point source emissions from limited number of atmospheric concentration measurements. Monache et al. (2008) applied Bayesian inference and Markov Chain Monte Carlo sampling to reconstruct a contaminant source on a continental scale. By running the Bayesian MCMC algorithm on a large parallel cluster the inversion results could be obtained in few hours as required for emergency response to continental-scale releases. Senocak (2008) developed a method for stochastic event reconstruction of atmospheric contaminant dispersion using Bayesian Inference. Cowles and Carlin (1996) wrote a review article describing the markov chain monte carlo convergence diagnostics. Convergence criteria to the right solution and optimum step size for chain movement are important topics in the application of Markov Chain Monte Carlo method. Haario et al. (2001) described the importance of choice of proposal distribution for MCMC methods for convergence of the algorithm. An adaptive metropolis algorithm with Gaussian proposal distribution was introduced. As shown in the above references, a number of researchers have worked in source location estimation using Bayesian probabilities methods. Since probability based methods alone cannot be sufficient during emergency response due to long convergence time periods, in the approach proposed in this paper, first an initial solution is quickly estimated using a least-square based method based on backward wind corridor. Subsequently, to obtain a more rigorous solution

GCPS 2013 __________________________________________________________________________

using Bayesian probabilistic method (referred as MCMC), the initial solution based on least square method is provided as a seed solution along with a search area/boundary. Thus, this twostage method decreases the overall convergence time for probabilistic method by providing an intelligent apriori information. Using this approach, the emergency responders can quickly start responding based on the initial solution and more accurate information using a probabilistic method can be provided to them in due course of time. Thus, a combined least square-Bayesian approach ensures that even a probabilistic method provides a solution in much smaller time frame.

2 Approach
In this section, our hybrid approach for estimation of source area location is described. At first, source area location is estimated using a simple and intelligent technique developed using backward wind corridor and least-square based method (Khajehnajafi and Burla, 2008). The least-square method is an optimization based approach and several key features of this approach are listed in Table. 1. This method is very efficient and gives a good estimate for source location for most of the wind stability conditions, except when the wind is too stable because the slope of wind corridor is very narrow. Another approach used for source location estimation is a probability based method using Bayesian theory where a probabilistic distribution is assumed for source area location and the sensor measurements and markov chain monte carlo (MCMC) simulations are used to estimate the source area location. Since this is a statistical approach, it takes significantly more time to converge to a value. Key features of this approach are listed in Table 1. A Hybrid approach is a combination of the two approaches as listed in Table 1. Deterministic approach assumes that the parameters of the system are known completely without any uncertainty while probabilistic approach accounts for uncertainty in the model at the expense of longer time for convergence. Therefore, having a method that can narrow down the initial search zone for the Bayesian approach could be very useful in decreasing the convergence time. Thus, a hybrid approach is suggested in this paper where at first a Source Area Locator (SAL) model is used to narrow down the area and then Bayesian approach (referred as MCMC) is used to zero down on the actual source area location. The idea is that the combination of these two is a great method that accounts for weaknesses of both the models and makes it a very efficient model to narrow down upon a location very close to the actual source for all wind and stability conditions. Each approach on its own can provide reasonable results with some limitation either in terms of a further away solution for certain atmospheric conditions, or really long convergence time. Thus, the hybrid approach is a very valuable approach to estimate the source area location. Next, specific details of each of the methods are provided.

GCPS 2013 __________________________________________________________________________

Table 1: Deterministic, Stochastic, and Hybrid Approach to Source Area Location


Deterministic (Least-square based) Stochastic (Probabilistic) Assumes that all the system parameters Assumes uncertainty in system parameters are known perfectly Usually longer run-time Usually faster runtime Probabilistic approach Good for emergency response situations Based on Bayesian - Markov Chain Monte Carlo method (MCMC) Referred as SAL Deterministic approach based on least-square based method to narrow down the search area Hybrid Approach Initially solved using deterministic approach based on least square method SAL Solution of deterministic approach is used to narrow down the search area for Bayesian Markov Chain Monte Carlo method (MCMC). Thus, reducing the run time and obtaining a more accurate prediction of source parameters.

2.1

SAL (Least Square Based Model)

The least square based method is referred to as SAL (Source Area Locator) in this paper. In this technique, it was demonstrated that with only the basic knowledge of meteorological information; wind speed, direction, and stability along with two or more gas concentration measurements, it is possible to quickly frame a region for the source of chemical (Khajehnajafi, and Burla 2008). A technique called reverse corridor or back trajectory is used to narrow the search domain for release location as seen in Figure 1 (a). This simple methodology uses the horizontal dispersion obtained from surface data at receptor locations. The construction of the corridor is key to reduce search area, therefore, reducing the computational cost. In Figure 1(a) the yellow region is the search domain if only sensor 1 and 2 were impacted. Introducing sensor 3 value, the search area is further reduced to the meshed zone. Adding additional sensors may narrow the search area; however, it will significantly increase the simulation time and increase the nonlinearity of the problem. For the simple example shown in Figure 1(a) initially, sensitivity of the model to main parameters for dispersion, e.g. wind meander, sensor concentration fluctuations, and atmospheric stability was studied. A Gaussian model was used for that study. The result of SAL model for different stability classes A through F (1 through 6) showed that the accuracy of the model prediction for source area decreased with the increasing in stability class especially for higher stability classes 5 and 6 and at very low release rates. The results for the source location and the release rates, even when the model was simulated with the noisy sensor data, viz. 40% and 80%, matched very well with the true data when very fine mesh size was use (except for highly stable conditions at low release rates) indicating that the back trajectory method coupled with a simple Gaussian plume model is in fact a very simple yet plausible technique for locating source of pollutant.

GCPS 2013 __________________________________________________________________________

Deterministic
Reverse Corridor Sensor 1
Reverse Search direction

Stochastic
Sensor 1

Sensor 2

Sensor 2

Sensor 3

Sensor 3

Wind Direction

Wind Direction

Search Area for Source Location

Search Area for Source Location

(a)
Figure 1:

(b)

Pictorial view of deterministic and stochastic approach to source location of hazardous material release

2.2

Bayesian Model (Markov Chain Monte Carlo)

As mentioned in the introduction section, MCMC alone takes a lot of time to converge. Herein, based on the solution from the SAL model, an area was outlined as a narrowed search domain for the Bayesian probabilistic method. If the initial guess is chosen at random, the outcome from MCMC may take much longer time to converge and may even fails to converge sometimes because of broader search zone. In Bayesian method, a prior distribution is defined for parameters of interest (For example the location and release rate in this case). The likelihood function is estimated using the predicted concentration and measured sensor concentration. Then both prior distribution and likelihood distribution are used to estimate the posterior distribution. The posterior distribution gives us the latest knowledge of the original parameters of interest. The Bayess theorem for the source reconstruction can be written as (Monache et al. 2008).

(1)

Where p( ) is the probability distribution. The x, y, q are the source coordinates and the release rate. Sensor data is the measurements from the different sensors. Bayes theorem relates the posterior distribution, , to the product of the probability of measurements given the source parameters, , also called the likelihood function, and the probability of the source parameters prior to any knowledge of the measurements, , also called the prior. For this study, x, y, q are assumed to be unknown parameters. One can include other parameters in the unknown list.

GCPS 2013 __________________________________________________________________________

2.2.1

Prior Distribution

Let m be the model parameters i.e. all the parameters for which prior need to be defined. x, y, q are the prior parameters, x coordinate of the release location, y coordinate of the release location and the release rate, respectively. The prior distribution can be written as:
(2)

2.2.2

Likelihood Function and Acceptance Condition

The likelihood distribution is function of CP, predicted concentration for a sensor, and CM, measured concentration by an actual sensor. Here, the likelihood function is assumed to be of the following form (Monache et al. 2008):
(3)

Where L is the likelihood function, CPi is the predicted concentration values at the sensor location i, CMi is the sensor measurements for sensor i, is an error parameter chosen accordingly to expected errors in the observations and predictions. N is the number of sensors. S is the state vector formed by all the unknown parameters (x, y, q). If the measured concentration values spread over several order of magnitudes, taking the logarithm of both predicted and measured values prevents large concentrations from dominating in the likelihood computation. The above likelihood is relatively simple, where the N prediction errors are assumed independent and expected to be proportional in size to the predicted concentration (as a result of log transforming both the observed and predicted concentration). If large errors are expected in the prediction/or measurements, large values of sigma should be assumed resulting in a broaderflatter posterior distribution. Here a sigma value of 0.8 is assumed. For a more complex scenario, it can also be treated as an unknown parameter and assigned a prior distribution. The posterior distribution is sampled with an MCMC procedure via a Metropolis-Hasting algorithm. 2.2.3 Concentration Prediction Model

The concentration prediction model is based on a Gaussian plume dispersion model as shown below (Kim et al. 2011).

(4)

GCPS 2013 __________________________________________________________________________

Where, is predicted concentration at a particular location (x, y, z). q is the emission rate, U is mean wind speed, H is height of the release, x is distance along the wind, y is distance along the horizontal crosswind direction, z is the distance along the vertical axis, are the standard deviations in the horizontal crosswind and vertical directions that are also called as Gaussian plume dispersion parameters and are defined empirically for different stability conditions. are defined below:
(5)

2.2.4

Sampling Procedure and Algorithm

Various algorithms exist for the sampling with MCMC. We have assumed uniform prior for the prior parameters. For the MCMC method, x is in the range -10,000 to 10,000 and y is the range 10,000 to 10,000 with (0,0) as the release location. q is uniform prior between 0.1 to 10 kg/s. U is the value given for the experiment of interest. H is assumed at the ground level. Ideally, parameter ranges can be assumed for wind direction, wind speed, 1, 2. However, for this study we have kept these constant. Of course, 1 and 2 will vary based on the stability class. For the hybrid approach, the release location is first determined using the SAL approach. Limits for x, y are assumed to be (-1000 + xsal, 1000 + xsal) and (-1000 + ysal, 1000+ ysal) where xsal and ysal are release location coordinates obtained using the least-square method (SAL). Range for q is 0.001 to 10 Kg/s. The new value of x, y, q is obtained using adaptive sampling based on the current value and a step size. 2.2.5 Algorithm

An initial value for each parameter is sampled from its prior distribution. System parameters are assumed to be uniform prior with range as described above. The prior distributions are not completely uncertain as a good initial range is provided using the SAL method. Using the likelihood function, these sampled values from the prior distribution are selected to be part of the posterior distribution based on some criteria. A number of iterations involving source term sampling and dispersion simulation are executed until convergence to the posterior distribution is reached. Accepted parameter values collected during all the iterations constitute a Markov Chain which is defined as a set of consecutive values of parameter in which the probability of a value depends only upon the previous sampled value. Each iteration of the solution that becomes part of Markov Chain consists of following steps (Monache et al. 2008): 1. New value for source term is proposed (xprop) according to xprop = x + dx. Where x is the current value and dx is displacement from the current value modeled as a random walk sampled from a Gaussian distribution with zero mean and a standard deviation specified from the current step size. The Prior distribution p (m) is used as a target distribution to estimate a prior likelihood of xprop. If the prior likelihood of xprop exceeds that of x, the proposed value xprop replaces x. If not, a random (Bernoulli) coin flip determines whether the new proposal, even with its lower prior likelihood will be accepted. This ensures that sampling doesnt get stuck at a random point.

GCPS 2013 __________________________________________________________________________

2. Repeat step 1 for the y and q parameters considered in the prior distribution. The assumption is that independent sampling of each parameter ensures that they reflect their likelihood better than sampling all of them at once and testing their likelihood 3. The predicted concentration for sensors is simulated using the current values of all the prior parameters using equation 4. 4. Likelihood of the current values of all the parameters which forms the proposed state is calculated using the predicted concentration and observed concentration by sensors. If the proposed likelihood is higher than the likelihood of previous state, it is accepted. If not, a random (Bernoulli) coin flip is used to ensure that the search reconstruction procedures continues to search the complete space of proposed source term parameter states, preventing the procedure from remaining indefinitely within the neighborhood of a local extreme. After the proposed state is calculated, it is accepted if
(6)

Where Lnew is the likelihood value of the proposed state and L is previous likelihood value and rnd(0,1) is a random number generated from a uniform distribution in the internal [0,1]. At a particular time instant, likelihood test for proposed state using measured data from all sensors at a particular time instant determines the acceptance or rejection of this state. Steps 1-4 are repeated to create a new proposed state with respect to all the parameters until the Markov Chain converges for a particular time instant. Note that in case of time varying information; step 1 to 4 can be further repeated for each time instant.

3 Case Studies: Comparison with the Field Data


3.1 BURRO EXPERIMENTS

BURRO experiments from the 1980 LNG (Liquefied Natural Gas) spill tests conducted by Lawrence Livermore National Laboratory (LLNL) and Naval Weapons Center (NWC) were used to compare the predictions from the model with the experiments. BURRO series experiments are one of the most comprehensive experimental analyses available in the literature for sensor measurement data of a chemical release (Koopman et al. 1980, Goldwire et al. 1980). In total nine experiments were conducted. Five experiments from the 1980 LNG (Liquefied Natural Gas) spill tests were used for this analysis. It is very difficult to get access to experimental data in this field and BURRO data are indeed very useful to compare modeling predictions with the experimental data for such a complex set of sensors. This particular experimental setup employs close to 20-30 sensors. Not all sensors will see concentration values all the time so only the sensor with non-zero concentration reading are analyzed in this situation. This in itself is debatable as non-zero sensors may provide some information about the location which otherwise can be missed in case of uncertainty. Ideally, non-zero sensors should be included as well but for simplicity reasons we have not included

GCPS 2013 __________________________________________________________________________

those here and can be part of a further study. The next level of complication arises from the dynamic nature of the measured sensor values due to plume meander caused by horizontal turbulent fluctuation in the atmosphere. In ideal situation one expects the sensors to measure concentrations reliably based on the release rates. However, the uncertainty in wind direction, wind speed, measurement errors due to hardware and software and variability in atmospheric stability condition can contribute to a lot of uncertainty in the sensor data. For example, see the BURRO experiments report for concentration values measured by sensors for few experiments. Because of these uncertainties in the values and the simplified nature of our modeling approach, careful and intelligent decisions need to be made about the measured sensor readings that are used for this analysis as only a single measured value is used for each sensor in this analysis. Nonetheless, a more complicated model can be created which can include few/all aforementioned uncertainties at the expense of higher simulation time and resources. These may be included in the future studies of the project after carefully analyzing the tradeoffs involved in various approaches. In our approach, the spill scenarios conditions and measured sensors values were used to obtain solution from SAL and MCMC approaches and then predictions for source area locations from the two models were compared against the actual experiment release location. In particular, five different spill scenarios from BURRO experimental series i.e. BURRO 2, BURRO 3, BURRO 7, BURRO 8, and BURRO 9 covering a range of atmosphere stability conditions are reported in this study (Table 2). The release location was denoted as (0, 0) position in a new coordinate system, which is about 45 degree from the true north. The wind conditions were measured along the true north conditions. Therefore, wind directions for the new coordinate system are calculated by adding 45 degrees to wind direction values reported in the BURRO experiments. A plot in the new coordinate system containing actual release location and sensor locations is shown in Figure 2. The sensors were mainly distributed in four different arches as shown in Figure 2. Detailed location of these sensors and other towers can be found in the BURRO report part 1. The wind fields for different experiments are reported in the BURRO report part 2. To run SAL, a constant wind field (i.e. constant wind speed and wind direction) is assumed. For these short duration experiments, wind field does not change a lot so this is a reasonable assumption. Undoubtedly including variability in the wind field would result in a better prediction at the expense of more computational time.

GCPS 2013 __________________________________________________________________________

Figure 2:

Burro experiment setup of sensors.

Table 2: Spill conditions for BURRO 3, BURRO 7, BURRO 8, and BURRO 9


Experiment Name BURRO 2 BURRO 3 BURRO 7 BURRO 8 BURRO 9 Chemical LNG LNG LNG LNG LNG Wind direction 266 270 253 280 277 Stability 2 2 3 5 4 Wind speed (m/s) 5.4 5.4 8.4 1.8 5.7 Spill Rate (m3/min) 11.9 12.2 13.6 16 18.4 Spill duration 173 166 174 107 79

3.2

SAL Sensitivity to Different Cost Functions

A reasonably good approximation for source area location using SAL is our first target. To accomplish that goal we had to identify the best cost function. Three cost functions were identified; A simple concentration difference objective function abs(cm-cp), referred as CF 1, Normalized concentration based objective function [abs(cm-cp)/cm ], referred as CF 2; Distance and concentration normalization based objective function [abs (cm-cp)/cm *(xdist2+ydist2)0.5], referred as CF 3. Where cm is the measured concentration and cp is the predicted concentration.

GCPS 2013 __________________________________________________________________________

Only a single measurement value per sensor is used for this analysis. Since it is a dynamic release scenario, various values can be selected from the sensor measurements. A sample of sensors output for Burro 3 is shown in Figure 3. Two set of data was used for SAL study. One was based on the single best sensor value for the entire time zone which reflects the best/average sensor values and the other is based on the maximum value for the entire time. All the sensors with non-zero sensor values were included in the analysis. For SAL simulation, thirteen different values of release rates ranging from 0.01 to 10 are used for source location. A local source point was identified which was average of these thirteen values and was used as an initial point for the next level search. First a coarse grid is used to narrow the search zone. Then a finer mesh grid search was carried out around that zone for various ranges of release rates to identify the source area location.

Figure 3:

Gas sensors concentration measurement at 140 m arc.

The following tables represent the single sensor value obtained from each sensor positioned in the experimental grids for this study along with the source location predicted by SAL using three different cost functions. 3.2.1 Burro 2

Table 3: Sensors value from BURRO 2 experiment


Sensors G2 G5 G4 G1 G8 G3 G12 G13 ppm 8 5 1.2 1 2 5 4 0.02

GCPS 2013 __________________________________________________________________________

Table 4: Source area location for BURRO 2 using sensor values


WD=266 Stability X location Y location CF1 2 -65.1 9.2 CF2 2 -57.5 7.5 CF3 2 -56.6 8.8

3.2.2

Burro 3

Table 5: Sensors value from BURRO 3 experiment


Sensors ppm G2 9.0 G5 4.0 G4 17 G1 15 G7 4 G6 5

Table 6: Source area location for BURRO 3 using sensor values


WD=270 Stability X location Y location CF 1 2 -73.4 -14.8 CF 2 2 -117.2 40.2 CF 3 2 -42.5 16.3

3.2.3

Burro 7

Table 7: Sensors value from BURRO 7 experiment


Sensors ppm G12 9.0 G3 0.78 T3 0.34 G13 2.8 G14 2.6 G18 0.4 G19 0.05

Table 8: Source area location for BURRO 7 using sensor values


WD = 253 Stability X location Y location CF1 3 -111.6 -49.6 CF2 3 69.5 59.0 CF3 3 191.6 148.6

3.2.4

Burro 8

Table 9: Sensors value from BURRO 8 experiment


Sensors ppm Sensors ppm G9 15 G12 2 G2 6 G10 0.02 T2 8 G16 0.7 G4 11 G15 0.7 G1 5 T5 2 G8 8 G17 3.8 G11 17 T6 0.25 G7 12.5 T4 3 G6 6 T3 5 G3 4.8

GCPS 2013 __________________________________________________________________________

Table 10: Source area location for BURRO 8 using sensor values
WD = 280 Stability X location Y location CF1 5 -633.9 119.8 CF2 5 -285.4 70.6 CF3 5 -355.2 77.9

3.2.5

Burro 9

Table 11: Sensors value from BURRO 9 experiment


Sensors G2 T2 G4 G7 T4 G6 T5 G16 T3 T6 ppm 6 1 7 1.2 2.2 6 0.25 0.2 4.5 0.2

Table 12: Source area location for BURRO 9 using sensor values
WD=277 Stability X location Y location CF1 4 -70 8.8 CF2 4 63.5 -3.9 CF3 4 66.3 -5.3

Figure 4:

Source location prediction for various Burro experiments using SAL (actual release is at 0,0).

GCPS 2013 __________________________________________________________________________

From this analysis we can see that the CF2 give better results out of all three objective functions. It was found that other than stability case 5, in most cases the source area was located within (+/70, +/- 70) range of the actual release location. SAL took only few seconds to run for all cases. 3.3 Hybrid Model

Subsequently, the solution from the hybrid approach with an initial guess from SAL to narrow down the search domain using Bayesian theory was obtained. BURRO MCMC refer to the results obtained using the hybrid approach in Figure 5. BURRO SAL refers to results obtained using the least-square based methodology. As evident from Figure 5, BURRO MCMC provides good solution for all stability conditions. Improvements in source location were seen for almost all the cases using the hybrid approach. These simulations were done for a wide variety of stability ranges including unstable to stable atmospheric conditions. Improvements seen in unstable cases using hybrid MCMC approach are more profound than stable cases as shown later in this report. The hybrid model improved the run time by a factor of two to three times.

Stability 5

y
Stability 2 Stability 3

x
Stability 3

Stability 4
Figure 5: Burro experiment result from SAL and hybrid MCMC approach

Time-varying wind field, uncertainty in wind direction and speed, dynamic sensor readings, release duration, release time may be included in prior parameters for the hybrid approach where the main focus will be on including uncertainties and then identifying the location of the release.

GCPS 2013 __________________________________________________________________________

If sensor data is available in dynamic fashion, these hybrid models can be extended to accommodate the dynamic Bayesian model (Johannesson et al. 2004) techniques.

4 Conclusion
The focus of this paper is mainly on estimating release location using the sensor measurements and providing an estimate for the release rate. An important method to estimate source area location using a hybrid approach based on least-square method (SAL) and bayeisan theory (MCMC) was provided. It was shown that the method provides a good solution for BURRO experiment data series for all stability conditions.

5 References
[1] [2] Tarantola A., Inverse Problem Theory and methods for model parameter estimation, Society for industrial and applied mathematics 2005, Philadelphia. Johannesson G., Hanley B., Nitao J., Dynamic Bayesian Models via Monte Carlo An Introduction with Examples -, UCRL-TR-207173, Lawrence Livermore National Laboratory, October 2004. Cowles M. K. and Carlin B.P. ,Markov Chain Monte Carlo Convergence Diagnostics: A Comparative Review, Journal of American statistical association, vol 91 (434), 1996, pp. 883. Yee E., Lien F-S, Keats A., DAmoursc R. , Bayesian inversion of concentration data: Source reconstruction in the adjoint representation of atmospheric diffusion, Journal of Wind Engineering and Industrial Aerodynamics, Vol 96, 2008, pp. 18051816. Yee E., Bayesian Probabilistic Approach for Inverse Source Determination From Limited and Noisy Chemical or Biological Sensor Concentration Measurements, Chemical and Biological Sensing VIII, Proc. of SPIE, Vol. 6554, 65540W, 2007. Kim J.Y., Jang H.K., and Lee J.K, Source Reconstruction of Unknown Model Parameters in Atmospheric Dispersion Using Dynamic Bayesian Inference, Progress in Nuclear Science and Technology, Vol. 1, 2011, pp. 460-463. Keats W.A. , Bayesian inference for source determination in the atmospheric environment, PhD thesis, University of Waterloo, Waterloo, Ontario, Canada, 2009. Neuman S., Glascoe L., Kosovic B., Dyer K., Hanley W., Nitao J., Event Reconstruction for atmospheric releases employing urban puff model UDM with stochastic inversion methodology, American Meteorological Society Conference, 2005, Atlanta GA. Zheng X., Chen Z., Inverse calculation approaches for source determination in hazardous chemical releases, Journal of Loss Prevention in the Process Industries 24 (2011) 293301. Sharan M., Singh S. K., and Issartel J.P, Least Square Data Assimilation for Identification of the Point Source Emissions, Pure and Applied Geophysics, August 2011.

[3]

[4]

[5]

[6]

[7] [8]

[9]

[10]

GCPS 2013 __________________________________________________________________________

[11]

Monache L.D. et al., Bayesian Inference and Markov Chain Monte Carlo Sampling to Reconstruct a Contaminant Source on a Continental Scale, Journal of Applied Meteorology and Climatology, pp 2600-2613, vol 47, 2008. Senocak I., Hengartner N. W., Short M.B., Daniel W.B., Stochastic Event Reconstruction of Atmospheric Contaminant Dispersion using Bayesian Inference, Atmospheric environment, Vol 42 (33), 2008, pp7718-7727. Koopman R. P., Baker J., Cederwall R. T., Goldwire H. C., Hogan W. J. LLNL/NWC (Lawrence Livermore National Laboratory/Naval Weapons Center) 1980 LNG (Liquefied Natural Gas) Spill Tests. Burro Series Data Report. Goldiwre H. C., Baker J., Koopman R. P., Cederwall R. T., Hogan W. J. LLNL/NWC (Lawrence Livermore National Laboratory/Naval Weapons Center) 1980 LNG (Liquefied Natural Gas) Spill Tests. Burro Series Data Report: The Appendices. S. Khajehnajafi, S. Burla, Prediction of Source Location and Rate of a Chemical Release for Emergency Response: Back Trajectory Approach, AIChE Spring Meeting, 2008 Hogan W.R., Cooper G.F., Wallstrom G.L., Wagner M. M. and Depinay J-M, The Bayesian aerosol release detector: An algorithm for detecting and characterizing outbreaks caused by an atmospheric release of Bacillus anthracis, Statist. Med. 2007; 26:52255252. Haario H., Saksman E., Tamminen J., An adaptive metropolis algorithm, Bernoulli, Vol 7 (2), 2011, pp. 223-242.

[12]

[13]

[14]

[15] [16]

[17]

Anda mungkin juga menyukai