Combined analysis of modeled and monitored SO₂ concentrations at a complex smelting facility

Abstract

Vale Canada Limited owns and operates a large nickel smelting facility located in Sudbury, Ontario. This is a complex facility with many sources of SO₂ emissions, including a mix of source types ranging from passive building roof vents to North America's tallest stack. In addition, as this facility performs batch operations, there is significant variability in the emission rates depending on the operations that are occurring. Although SO₂ emission rates for many of the sources have been measured by source testing, the reliability of these emission rates has not been tested from a dispersion modeling perspective. This facility is a significant source of SO₂ in the local region, making it critical that when modeling the emissions from this facility for regulatory or other purposes, that the resulting concentrations are representative of what would actually be measured or otherwise observed. To assess the accuracy of the modeling, a detailed analysis of modeled and monitored data for SO₂ at the facility was performed. A mobile SO₂ monitor sampled at five locations downwind of different source groups for different wind directions resulting in a total of 168 hr of valid data that could be used for the modeled to monitored results comparison. The facility was modeled in AERMOD (American Meteorological Society/U.S. Environmental Protection Agency Regulatory Model) using site-specific meteorological data such that the modeled periods coincided with the same times as the monitored events. In addition, great effort was invested into estimating the actual SO₂ emission rates that would likely be occurring during each of the monitoring events. SO₂ concentrations were modeled for receptors around each monitoring location so that the modeled data could be directly compared with the monitored data. The modeled and monitored concentrations were compared and showed that there were no systematic biases in the modeled concentrations.

Implications:

This paper is a case study of a Combined Analysis of Modelled and Monitored Data (CAMM), which is an approach promulgated within air quality regulations in the Province of Ontario, Canada. Although combining dispersion models and monitoring data to estimate or refine estimates of source emission rates is not a new technique, this study shows how, with a high degree of rigor in the design of the monitoring and filtering of the data, it can be applied to a large industrial facility, with a variety of emission sources. The comparison of modeled and monitored SO₂ concentrations in this case study also provides an illustration of the AERMOD model performance for a large industrial complex with many sources, at short time scales in comparison with monitored data. Overall, this analysis demonstrated that the AERMOD model performed well.

Introduction

Air pollution dispersion modeling and air quality monitoring are two methods of determining pollutant concentrations that can be used to gauge the impact a facility has on local air quality and to evaluate compliance in many regulatory jurisdictions. Monitoring can provide accurate concentration measurements at specific locations; however, monitoring can be costly and can only be done at a limited number of locations. Modeling can be used to assess concentrations anywhere; however, the quality of the modeled results is dependent on the accuracy of the model itself and on the reliability of the inputs. When data from both methods are combined and analyzed together, the accuracy of the modeling can be assessed and can yield a more comprehensive picture of the typical emissions from a facility and their effect on the local community. In Ontario, the use of such analysis is known as a Combined Analysis of Modelled and Monitored Data (CAMM) and is written into Ontario regulations (Ontario Regulation 419/05) as a method of ensuring that the emission rates used in a dispersion model for regulatory purposes are as accurate as possible (highest possible data quality).

Unlike a typical dispersion model evaluation study that would involve monitoring and modeling emissions from a well-defined source with well-understood emission rates and flow parameters, the CAMM process often involves multiple emission sources, some of whose emission rates (and possibly also flow parameters) have not been accurately measured, and instead have been estimated with a degree of uncertainty. The monitoring is designed to isolate the contributions of those sources whose parameters are less well understood, and identify and correct dispersion model biases that may be due to poorly estimated source parameters, especially emission rates. The CAMM is akin to other approaches that use a combination of monitoring and dispersion modeling to estimate an unknown emission rate for a source, but in the present case, it involved an industrial complex with a variety of emission sources in proximity to each other. This introduces a higher degree of complexity and requires rigor in the design of the monitoring data, filtering of the data, and assessment of the likely causes of observed model biases.

The objective of a CAMM is to identify and moderate any systematic biases in modeled concentrations based on comparison with actual measured (monitored) concentrations. Biases in the model could be due to a number of factors, including the model itself, meteorological inputs, stack and building characteristics, and uncertainties or omissions in emission data. Successful moderation of these biases results in a more accurate model with improved representation of the facility's impact on local air quality. As part of the CAMM process, significant effort goes into ensuring that the best available meteorological data, terrain data, stack parameters, and building dimensions are incorporated into the dispersion model. The assumption is then made that these data sets do not introduce any systematic bias. It is also assumed that the dispersion model itself is unbiased. It is well known that dispersion models generally do not compare well with monitoring data, when the model and monitoring data are paired in time and space (Hanna et al., 1999; Weil et al., 1992). Nevertheless, the models can perform with sufficiently little systematic bias (less than a factor of 2) to provide meaningful results in a statistical sense. The final key assumption of the CAMM process is that any observed model bias, or at least any bias beyond what might normally be expected from a dispersion model, is due to inaccurate estimates of emission rates for those sources whose emissions are not well understood. The observed bias can then be corrected by adjusting the estimated emission rates.

The subject of this study was a large nickel smelting facility owned and operated by Vale Canada Limited (Vale) located in Sudbury, Ontario. This is a complex facility with many sources of SO₂ emissions, including a mix of source types ranging from passive building roof vents to North America's tallest stack. In addition, as this facility performs batch operations, there is significant variability in the emission rates depending on the operations that are occurring. Although SO₂ emission rates for many of the sources have been measured by source testing, the reliability of these emission rates has not been tested from a dispersion modeling perspective. Source tested emission rates are for worst-case operating scenarios of each process, but it is not known how the overall SO₂ emissions from the facility vary on a day-to-day basis.

This facility is a significant source of SO₂ in the local region, making it critical that when modeling the emissions from this facility for regulatory or other purposes, the resulting concentrations are representative of what would actually be measured or otherwise observed. To assess the accuracy of the modeling, a detailed analysis of modeled and monitored data for SO₂ at the facility was performed.

Although the purpose of the present paper is to demonstrate the CAMM process for regulatory purposes, this type of study also provides the opportunity to assess the reliability of the dispersion model used, in this case AERMOD (American Meteorological Society/U.S. Environmental Protection Agency Regulatory Model). There have been a number of other field studies that were used to evaluate AERMOD against other dispersion models (Perry et al., 2004). These studies covered a variety of conditions and AERMOD was generally shown to perform well in comparison with other dispersion models. The facilities tested, however, were relatively simple with a limited number of sources and source types and with well-understood source parameters and emission rates. The present study allows for an evaluation of AERMOD at a complex facility with approximately 80 sources of SO₂ and from a range of source types with highly variable exhaust parameters and estimated emission rates of varying accuracy.

Methodology

Monitoring program

For this study, a mobile monitoring campaign measured SO₂ concentrations at five preselected locations in relatively close proximity to the sources at the smelter complex within the Vale property. These locations were selected such that they would measure SO₂ concentrations predominantly from sources that

1.	account for significant proportions of the SO2 emissions;
2.	were indicated by previous modeling to account for significant ground level concentrations of SO2; and
3.	have uncertainty in the SO2 emission rates related to the batch process operations.

Furthermore, the monitoring locations were selected in an attempt to isolate the contribution from individual groups of related sources (source groups) as much as possible.

To inform the selection process for monitor locations, preliminary AERMOD modeling was used to predict the regions where a relatively high potential for elevated concentrations would be observed under different wind directions. This information was used to determine where among the five sites the mobile unit would be stationed under a given wind condition. The preliminary modeling was also used to investigate locations where contribution from different source groups could be isolated as much as possible, which would assist in identifying the source or group of sources responsible for any observed model bias. The selected monitoring locations are shown in Figure 1 along with all source locations. All monitoring locations were approximately 300–500 m away from the dominant sources of interest.

Figure 1. Smelter complex sources and monitoring locations.

Real-time monitoring was carried out with a Teco 43i SO₂ pulsed fluorescence gas analyzer (Thermo Scientific). The minimum detection limit of this monitor is 2 ppb, with a maximum range extending to 100 ppm. This analyzer records concentrations every minute, and the values are averaged to provide 5-, 10-, 30-, and 60-min average values. The monitor was mounted onto the back of a mobile vehicle so that it could easily be moved to different monitoring locations depending on the wind directions.

As outlined in Standard Operating Procedure for SO₂ provided in the Operations Manual for Air Quality Monitoring in Ontario, an internal zero/span check was performed at the Vale garage and reviewed on a daily basis (Ontario Ministry of the Environment, 2008). This check lasted approximately 20 min and was performed before and after each day of monitoring. Additional calibrations using a certified calibration unit were performed at the start and the end of the program.

Monitoring was done during weekday daytime hours from July through November 2011. Current and forecast wind directions for the day were obtained in the morning and any changes throughout the day were monitored such that an appropriate sampling location could be selected and changed if and when required due to shifting winds. Attempts were made to keep the monitor in the same location for a minimum of one clock-hour to coincide with a modeled hour. However, there were instances where the operator aborted the scan if the readings approached zero for a significant portion of the hour, indicating an unexpected wind or operational change. Coordinates of the mobile monitor location were recorded using a Global Positioning System (GPS) unit to ensure the same locations would be used in the model.

The duration of the monitored period was dependant on whether or not non-zero SO₂ concentrations continued to be read from the equipment. No measurements were taken during periods of precipitation or fog.

Review and screening of monitoring data

The monitored data from July through November were reviewed, and periods of detected SO₂ concentration that lasted less than 45 min were discarded as they were deemed to be too short to be representative of conditions occurring over a full hour. In general, this occurred only if measured concentrations dropped to zero due to a shift in the wind direction. The monitoring campaign resulted in 255 “hits,” where a hit is defined as a single monitoring period that lasted a minimum of 45 min and up to 1 hr. Hits that occurred consecutively at the same location were averaged to become a single “event.” An event is defined as one or more consecutive hits at the same monitoring location where the concentrations for the individual hourly hits were later averaged to represent a single data point in the CAMM analysis. Averaging over consecutive hits was thought to provide a more robust, smoothed datum from both the monitoring and the corresponding dispersion modeling for the event, compared with using the individual 1-hr hits.

Once this initial analysis of the monitoring data was completed, the data were then scrutinized to ensure that only the data that would provide the most insight into the model performance would be used for the later analysis. This process ensures that only the most reliable data are used if adjustments to emission rates are found to be required. The criteria used to screen out monitoring data were

1.	missing meteorological data;
2.	low wind speeds;
3.	monitor not being downwind of smelter complex SO2 sources due to wind shifts during a monitoring event; and
4.	low monitored concentrations.

There were 3 days of monitoring where pertinent meteorological data were not available for the dispersion modeling analysis (described below) of the monitored period; thus, it was necessary to screen out monitoring events occurring on August 15 and 16, and September 12, 2011.

In accordance with Ontario CAMM Guidance, monitored hits where the wind speed was less than 2 m/sec were also screened out. Dispersion during periods of calm winds is less predictable by Gaussian models and these calms are often associated with variable wind directions (Ontario Ministry of the Environment, 2011). For this reason, both the modeled and monitored data for these events are less reliable and were thus screened out.

Although attempts were made to always monitor downwind of the smelter sources, there were a few instances where this was not the case for at least a portion of a monitoring event. However, hours were only screened out if the wind was not blowing directly from any of the smelter complex SO₂ sources for at least 50% of the event. In other words, if the wind was blowing from a different wind direction for only 1 hr of a 4-hr event, this period remained as a 4-hr event and the data were considered valid.

After a review of meteorological conditions, measured SO₂ concentrations for each event were averaged. The final screening step was then to remove events where the averaged concentration was less than 5 ppb (˜13 μg/m³). This value was selected so as to only represent monitored concentrations above typical background levels.

The screening analysis resulted in 168 hits, or a total of 77 events. Table 1 provides a summary of these data.

Table 1. Summary of events

Monitoring Location	Initial Number of Hits^a	Number of Hits Screened Out	Final Number of Hits	Final Number of Events Used (Consecutive Hits Combined)
A	93	35	58	25
B	41	25	16	6
C	60	12	48	25
D	35	5	30	14
E	26	10	16	7
All locations	255	87	168	77
Note: ^aHit is defined as a single monitoring period lasting between 45 min and 1 hr.

Dispersion modeling

Dispersion modeling of the monitored events was performed using the U.S. Environmental Protection Agency (EPA) AERMOD dispersion model version 11103 and the associated Building Profile Program Plume Rise Model Enhancements (BPIP-PRIME) application for modeling building downwash effects. Although three new versions of AERMOD (versions 11353, 12060, and 12345) have been released since this study has taken place, none of the changes is expected to have a significant effect on this analysis. The most significant change in recent versions was the incorporation of more flexible modeling options for low wind speed events in version 12345; however low wind speed events have been screened out from this analysis as mentioned above and thus these options are not applicable. A test has shown that the results for dispersion modeling of the facility with AERMOD version 12345 and with AERMOD version 11103 are very similar.

Each individual monitoring hit was modeled as a single hour. Similar to the way consecutive monitoring hits were averaged to obtain a concentration for a single event, concentrations obtained from the modeled hits were averaged for the same consecutive hours.

All SO₂ sources at the Vale smelter complex are well known and were included in the modeling. The source parameters such as exhaust temperatures and velocities were obtained from either source testing measurements for the most significant sources, or from design ratings or engineering judgment for some of the less significant sources. The modeled emission rates varied depending on the production activities that were occurring during each monitored hit, as is described below. All buildings were also included in the modeling to account for building downwash effects.

Since the goal of a CAMM is to compare modeled and monitored data to detect any bias, each event must be modeled to match the conditions of the monitoring as accurately as possible. This includes using site-specific meteorological data for the same hour(s) as the monitoring, using receptors at the same location as where the monitor was located, and ensuring that the emission rates used represent those experienced during the monitored event.

The site-specific meteorological surface and profile files were processed using surface data from Sudbury Airport and upper air data from the Detroit (formerly Whitelake) Upper Air Station and site-specific land use parameters.

For each modeled hit, a receptor was placed at the same location where the monitoring occurred using the coordinates obtained with the GPS unit, and at a height of 1.5 m above grade to correspond with the approximate inlet height of the monitor. Four additional receptors were placed 50 m to the north, south, east, and west of the central receptor. These additional receptor locations were chosen in an effort to reduce the impacts of meandering winds or other discrepancies between the actual wind direction transporting the facility emissions and the wind directions in the meteorological data set. The concentrations modeled at the five receptor locations were later averaged for comparison with the monitored concentration.

As this facility performs batch operations, there is significant variability in the various emission rates depending on what production activities are occurring. Great effort was invested into estimating the actual emission rates from each source in attempt to represent the production activities occurring during each of the monitoring events. Vale maintains detailed records with regards to when operations were occurring for each process as well as various parameters such as material flows, etc. Therefore, past source testing was compared with the specific operations that were occurring at that time to find a correlation that links production levels to SO₂ emission rates. This correlation was then used to estimate the emission rate occurring during each monitored hit based on the production at that time. This process was repeated for each source.

Modeled and monitored data analysis

A number of analysis techniques were used to compare the modeled and monitored data and assess if there is any bias in the modeled concentrations.

First, a paired analysis was performed where the modeled concentrations are graphed along with their corresponding monitored concentrations. This was analyzed qualitatively to provide a means with which to easily identify specific events where the modeled and monitored results are in strong agreement or disagreement.

To provide an indication of any overall bias, quantile-quantile (q-q) plots were produced. This is an unpaired concentration plot, where the data sets are ranked and the values of corresponding ranks are plotted as a data point. In other words, the highest modeled value is plotted versus the highest monitored value, the second highest modeled value is plotted versus the second highest monitored value, etc. This type of plot emphasizes any significant systematic bias present in the results. For example, if the unpaired results are consistently above the 1:1 line, then there is a strong bias toward modeling (assuming modeled values are the ordinate).

Lastly, the fractional bias (FB) and the robust highest concentration (RHC) were used as metrics to quantitatively assess whether or not there is any bias in the model. The FB is calculated as

(1)

where μMo and μMd are the arithmetic averages of the monitored and modeled concentration data sets. A value of zero indicates that there is no fractional bias, a positive value indicates that there is a bias towards the monitored concentrations, and a negative value indicates that there is a bias towards the modeled concentrations.

The RHC is a method of calculating the highest modeled and monitored concentrations based on information contained in the upper end of the distribution (Cox and Tikvart, 1990). This way the RHCs are less sensitive to unusual events in comparison with the actual highest concentrations. The RHCs are calculated as follows:

(2)

where R is the number of values used to characterize the upper end of the concentration distribution (taken here as 26, as recommended by Cox and Tikvart, 1990), X{R} is the Rth highest concentration, and X _bar is the geometric mean of the R − 1 largest values. If the ratio of the modeled RHC to the monitored RHC is equal to 1, then there is no bias in the model with respect to this method. A ratio that is greater than 1 indicates a bias towards modeling and a ratio that is less than 1 indicates a bias towards monitoring.

Site plan plots were also created for each event that visually display the facility along with all sources, the modeled receptor locations, and a wind rose corresponding to the hour(s) of the event. Also listed on these figures are the overall modeled and monitored concentrations, and the modeled emission rates and resulting concentrations of each individual source group (groups of sources from the same process). These plots are extremely valuable to visually observe how the wind was blowing during each particular event and therefore what sources were expected to be contributing to the monitored concentration. Furthermore, the modeled concentrations of each individual source group can provide an indication of the possible sources most likely to be responsible for any modeling bias for each particular event.

Regardless of whether it is determined that there is an overall modeling bias or not, it is important to study the data provided in the site plan plots thoroughly. For example, these data may suggest that the emission rates from one source group are consistently being overestimated, whereas the emission rates from another source group are consistently being underestimated. The end result will be that the q-q plot will show that there is no overall bias, and yet the site plan plots can help to identify situations where there may be a bias in the emission rates from individual source groups.

Results and Discussion

Figure 2 shows the results of the paired comparison between the modeled and monitored concentrations at each monitoring location. Note that each bar may represent more than one consecutive hour of data that were averaged. Events joined by a comma such as “A45,A46” were initially considered separate monitoring events because there was a pause of approximately 5 min between events. However, since the pauses were very brief in duration, the events were later combined.

Figure 2. Paired modeled and monitored SO₂ concentrations at (a) Location A, (b) Location B, (c) Location C, (d) Location D, and (e) Location E.

Modeled concentrations appear to be within a factor of 2 of the monitored concentration for most events; however, there are still some significant discrepancies between modeled and monitored concentrations for certain individual events. It was previously noted that dispersion models do not compare well with monitoring data when the data are paired in space and time. Therefore, although great effort was invested in constructing meteorological data sets, emission rate estimates, and other model inputs to be as representative as possible for each specific event, some variation between measured and monitored values is still expected.

In terms of the meteorological data, random differences between the local meteorological conditions and those at the meteorological station that provided input data for the dispersion model may result in over- or underpredicted concentrations for individual events. However, this uncertainty is not expected to affect the overall bias, since as long as the data set is large enough, these random differences will average out. Thus, any remaining bias can be attributed to the modeled emission rates.

Figure 3 shows the unpaired q-q plot of the modeled versus monitored concentrations, which give an indication of model bias. In addition to showing the modeled and monitored data points, the three solid lines represent the 1:1 line (where the modeled and monitored concentrations would be equal), and the 2:1 and 1:2 lines (where the modeled concentrations would be a factor of 2 higher or lower, respectively, than the monitored concentrations).

Figure 3. Unpaired q-q plot of the modeled versus monitored SO₂ concentrations.

This figure demonstrates there is no significant bias in the model, as all data points fall within the 1:2 and 2:1 lines, and follows the 1:1 line fairly closely for the higher concentrations. From a regulatory perspective, it is the higher concentrations that are the most important to be unbiased, since these are the values that are generally used to assess whether the facility is in compliance with applicable regulations. In this case, the highest quarter of the data shows little to no bias overall.

The data set had an overall fractional bias of −0.07, indicating that there is only a very slight bias towards the modeled concentrations (i.e., the modeled concentrations averaged 7% higher than the monitored values). From a regulatory perspective, this is favorable because it indicates that the model may be slightly conservative. The ratio of the modeled to monitored RHCs, however, was calculated to be 0.93, which shows a slight bias towards the monitoring data when considering the upper end of the distribution. These metrics demonstrate that overall the biases in the modeling were minor.

Part of the reason that this analysis has shown little overall bias is likely due to the rigorous screening approach that has been applied as was described above. Again, this is an essential step in the CAMM process to ensure that only the most reliable data were used.

Lastly, site plan plots provide more detailed information for individual events. These were created and analyzed for each event, but an example of such a plot is shown in Figure 4.

Figure 4. Example of a site plan plot for Event A1.

For this particular event, the site plan plot shows that the modeled concentration is overestimating the monitored concentration. If it is assumed that the model itself and the meteorology are accurate, then for this particular event, it is likely that the emission rates from the Weak Acid Treatment Process Stack and the MPV Roof Monitors are being overestimated, since they are the main contributors to the high modeled concentrations. However, in studying the site plan plots for each event (not shown), there were no trends of source groups that consistently appeared to be resulting in over- or underestimated modeled concentrations. Therefore, it was also concluded that there was no bias from any individual source groups.

Overall, since no significant bias was observed, and the site plan plots did not show any additional bias for specific source groups, the emission rates used were considered representative and the facility was modeled appropriately. Had a significant bias been observed, further work would have been done to identify the source or group of sources that was most likely responsible for the bias. This would have involved a more careful analysis of the data to see which of the five monitoring location(s) appeared to contribute most to the overall bias. Since the monitoring stations were selected so as to isolate particular sources or source groups, this would have given an indication of which emission source(s) or source group(s) was most responsible for the observed model bias. In addition, since the source emission rates were highly variable, observing which sources were operating during the events for which the bias was most significant would provide an indication of the likely responsible source(s). The source(s) found to be responsible for the bias would require adjustments to the emission rates. This was not necessary in the present case study, however, as the model proved to be relatively unbiased without any adjustment of the initial emission estimates. It is believed that this result can be credited to the amount of initial effort used to ensure that each event was modeled as accurately as possible to represent the true conditions occurring during that time.

Summary

A CAMM analysis for SO₂ was undertaken for Vale's Copper Cliff smelter complex. This involved the use of monitoring data in combination with dispersion modeling to refine the emission estimates for sources whose emission rates were not accurately known. The smelter complex is large with a wide variety of emissions sources, which significantly increased the amount of care and effort needed in designing the monitoring program and filtering the data. The CAMM included collecting on-site monitored SO₂ concentrations at a number of locations, screening of the collected data to obtain only representative concentrations, a detailed review of the smelter complex sources and production data to determine variable emission rates for use in the modeling, and comparison of the modeled concentrations to the monitored data for the 77 identified events. The results of the CAMM showed that overall there was no significant bias in the modeled concentrations. Therefore, the emission rates used were representative and the facility was modeled appropriately. It is believed that this result can be credited to the amount of initial effort used to ensure that each event was modeled as accurately as possible to represent the true conditions occurring during that time.