



Date:	February 7, 2013

Subject: 	Technical Approach for Wool Fiberglass MACT Floor Calculations
	40 CFR 63, Subpart NNN  -  Supplemental Proposal
	EPA Contract No. EP-D-11-084; EPA Work Assignment No. 1-07
	RTI Project No. 0213199.001.007

From: 		Cindy Hancy
		Dave Reeves

To: 		Susan Fairchild, EPA/OAQPS/SPPD/MMG

Introduction
Section 112 of the Clean Air Act (CAA) requires that the U.S. Environmental Protection Agency (EPA) establish National Emission Standards for Hazardous Air Pollutants (NESHAP) for the control of the hazardous air pollutants (HAP) emitted from both new and existing major sources in a source category.  These standards must reflect the maximum degree of reduction in HAP emissions that is achievable.  The minimum level of control is referred to as the "Maximum Achievable Control Technology (MACT) floor."  The method for determining the MACT floor for a NESHAP is defined for both new and existing sources in CAA section 112(d)(3).  For new sources, the MACT floor cannot be less stringent than the emission control that is achieved in practice by the best-controlled similar source.  For existing sources, the MACT floor cannot be less stringent than the average emission limitation achieved by the best-performing 12 percent of existing sources for source categories with 30 or more sources, or the best-performing five sources for source categories with fewer than 30 sources.
The purpose of this memorandum is to present the data, methodology, and results of the MACT floor analysis for the Wool Fiberglass Manufacturing NESHAP, 40 CFR 63, Subpart NNN source category. This analysis is part of EPA's obligation under CAA section 112(f)(2) and 112(d)(6) to conduct a residual risk and technology review.  The MACT floor analysis is also being performed in response to a petition for rulemaking by the Natural Resources Defense Council and Sierra Club that states EPA failed to set emission limits for HAPs (phenol and methanol) emitted by wool fiberglass facilities and challenges some HAP surrogacies used in the current wool fiberglass NESHAP. This MACT floor analysis uses data collected from a nationwide voluntary survey of wool fiberglass manufacturers conducted by EPA in 2010 and an Information Collection Request (ICR) sent out to the wool fiberglass industry in November 2011.  Data on process operations, emission controls, and air emissions data reported by respondents to the survey and ICR were compiled into a Microsoft Access data base that serves as the data set used for this MACT floor analysis (referred to in this memorandum as the "ICR data set").
      
On November 25, 2011 EPA proposed revisions to the Wool Fiberglass Manufacturing NESHAP, 40 CFR part 63, subpart NNN.  This memorandum incorporates changes made since the November 2011 proposal to reflect public comments regarding facility information and emissions data.

Background Information
The current Wool Fiberglass Production NESHAP (subpart NNN) was promulgated on June 14, 1999 and applies to each of the following existing and newly constructed major sources located at a wool fiberglass manufacturing facility: glass-melting furnaces, rotary spin (RS) manufacturing lines that produce bonded building insulation, and flame attenuation (FA) manufacturing lines producing bonded pipe insulation. The current rule also applies to new FA manufacturing lines producing bonded heavy-density and pipe products.  Currently, subpart NNN defines "bonded" as wool fiberglass to which a phenol-formaldehyde binder has been applied. The use of these types of HAP binders results in formaldehyde, phenol, and methanol emissions from sources at bonded RS and FA lines which generally consist of forming, curing, and collection operations.  RS and FA manufacturing lines that produce nonbonded products are not currently subject to the standards.   The 1999 Wool Fiberglass Production NESHAP (subpart NNN) sets particulate matter (PM) emission limits for new and existing furnaces and sets formaldehyde limits for new and existing RS lines and for new FA manufacturing lines.  The original NESHAP also uses the surrogate approach, where PM serves as a surrogate for HAP metals and formaldehyde serves as a surrogate for organic HAP. 

The wool fiberglass manufacturing source category currently consists of 30 facilities operating in the United States; 10 of these facilities are major sources and operate 29 glass-melting furnaces. The remaining 20 facilities are listed as area sources or synthetic minors. The North American Insulation Manufacturers Association (NAIMA) announced that by the end of 2012, they predict no major sources subject to Subpart NNN will exist in this source category (see docket items 0040 and 0052) by phasing out the use of phenol-formaldehyde binders. Once a facility has phased out its use of these HAP binders on bonded lines, it may apply to the appropriate permitting authority to no longer be subject to the MACT Subpart NNN standard under two definitions of the MACT rule: the applicability definition and the definition of `bonded'. 
      
The specific chemicals, compounds, or groups of compounds designated as HAP are listed in CAA section 112(b). From this list, chromium compounds, hydrogen fluoride (HF) and hydrogen chloride (HCl) were identified as being emitted from furnaces; and formaldehyde, phenol, and methanol were identified as being emitted from bonded FA and RS lines.  In the November 2011 proposal, EPA set MACT floor limits for HF and HCl. The Agency received comments on the proposed HF and HCl emission limits and upon reexamination of our analysis of the acid gas test data, we found that up to 80 percent of the HF and HCl test data in the floor were below detection limit (BDL) of the required test method. Based on this information, EPA has decided to set work practice standards for HF and HCl instead of MACT floor emission limits under section 112(h) of the CAA. The 2011 proposal also set emission limits for chromium compounds based on both section 112(d)(6) technology review and 112(f)(2) residual risk ample margin of safety. The Agency is not revising the chromium emission limit in the 2011 proposal, but is correcting the basis for proposed limit being technology review (and not a calculated MACT floor). Details regarding chromium compound emission limits and particulate matter (PM) limits for glass-melting furnaces are discussed in the Section 112(d)(6) Technology Review for Wool Fiberglass NESHAP memorandum and in the Draft Risk Report for the Wool Fiberglass Manufacturing RTR available in the docket for this rulemaking (docket # EPA-HQ-OAR-2010-1042). With regards to formaldehyde, phenol, and methanol at RS and FA bonded lines, EPA proposed to remove the surrogacy relationship in the rule and establish limits for phenol and methanol and correct the formaldehyde limit based on data submitted by the industry in their voluntary information collection effort of 2010. After the 2011 proposal, EPA received information from industry clarifying that the proposed limits for formaldehyde, phenol and methanol were based on a data set which included test data for lines that had phased out use of phenol/formaldehyde binders and were therefore no longer subject to the rule. The Agency removed the erroneous data from the dataset and recalculated the MACT floor for RS lines based on the bonded lines at major sources and synthetic minor sources. The limits in the supplemental proposal for RS lines differ from the proposed limits due to the removal of erroneous data from the data set used to calculate the MACT floor.
MACT Floor Analysis - Subcategories
Under CAA section 112(d)(1), EPA has the discretion to "...distinguish among classes, types, and sizes of sources within a category or subcategory in establishing..." standards.  When separate subcategories are established, a MACT floor is determined separately for each subcategory.  To determine whether the wool fiberglass manufacturing facilities warrant subcategorization for the MACT floor analysis, EPA reviewed unit and process designs, operating information, and air emissions data compiled in the ICR data set and other information collected by the Agency for development of the NESHAP for this source category. The EPA also analyzed several subcategorization recommendations submitted by industry in their public comments on the proposed rule (76 FR 72770, November 25, 2011). Based on this review, EPA concluded that sufficient information exists to continue subcategorization of bonded lines into both FA and RS lines, but that beyond this level, there are no significant design or operational differences at wool fiberglass facilities that warrant further subcategorization. Specifically, the original/current wool fiberglass NESHAP (subpart NNN) contains separate emission limits for FA manufacturing lines that produce heavy-density wool fiberglass and FA manufacturing lines that produce pipe product wool fiberglass. According to recent industry information, there is only one company that runs FA lines at their facilities. These facilities provided test data in response to the voluntary survey that was to represent all FA lines. The EPA received data for three  FA lines that were running heavy-density product and no data on FA lines that were running pipe products. In their May 5, 2010 letter to EPA (see Docket EPA-OAR-HQ-2010-1042), industry stated that the testing at pot & marble flame attenuated (FA) process units was representative of all products made on FA lines based on the following: "all re-melt borosilicate glass marbles; all are pot & marble flame attenuation; and all are operated by Johns Manville."  The letter also stated that Johns Manville is the only NAIMA member with a pot & marble melting process and the FA process is "a completely different process than rotary (spin) and was subcategorized in the original MACT standard." We therefore concluded that the limits developed for FA lines were representative of all products made on FA lines and that further subcategorization was no longer supportable under the standard. We proposed only limits for FA lines without regard to further subcategorization, and all products made on an FA line using phenol/formaldehyde binders would be subject to those limits.  Although we received comment suggesting subcategorization of up to four subcategories on FA lines, we received no additional test data from industry to support further subcategorization beyond the general FA line subcategory.  Consequently, we propose limits for FA lines such that all products made on an FA line using phenol/formaldehyde binders are subject to the proposed emission limits.
MACT Floor Analysis Methodology
Existing Sources 
A MACT floor analysis was completed for formaldehyde, phenol, and methanol emissions from RS and FA lines. The majority of the data (e.g., production and equipment throughput rates) used to calculate the MACT floors was submitted as confidential business information (CBI) and therefore, the data and calculations are not publicly available (and not included in this memorandum). 

The first step in the MACT floor analysis for each regulated source and HAP was to array all the data and rank each unit (for which emissions data was provided) by emission level (lowest to highest) for each pollutant.  From this ranking, a MACT floor pool of sources was identified for determining the minimum control level allowed for the MACT floor, consistent with the criteria defined for new and existing sources by CAA section 112(d)(3).  For the new source MACT floors, the best-controlled source was identified for which there were individual source test run data in the ICR data set.  For the existing source MACT floors, selection of the MACT floor pool size (i.e., number of emission units to be included in the determination of the average emission limitation value) was determined on an individual unit category basis as described below. 
      Bonded RS Manufacturing Lines.  There are less than 30 RS lines located at the 10 major sources; therefore the top 5 best performing RS lines for which emission test data was provided were used in calculating the MACT floor for formaldehyde, phenol, and methanol. For new sources, the best performing source was used to calculate the MACT floor for each pollutant.  
      Bonded FA Manufacturing Lines.  This category includes less than 30 sources and emission test data was submitted for three FA lines. Therefore, all of these three FA lines were used to calculate formaldehyde, phenol, and methanol limits for existing sources. For new sources, the best performing source was used to calculate the MACT floor for each pollutant.  
The next step in the MACT floor analysis was to account for data variability in the calculations of the applicable MACT floor limits for the subcategories using the data's 99% upper prediction limit (UPL).  Specifically, the MACT floor limit was determined as the UPL calculated with the Student's t-test using the "TINV" function in Microsoft Excel software.  The UPL approach has also been used in other EPA rulemakings (e.g., NESHAP for Portland Cement, NSPS for Hospital/Medical/Infectious Waste Incinerators, NESHAP for Industrial, Commercial, Institutional Boilers and Process Heaters, and NESHAP for Electric Generating Units) to account for variability in emissions data for a specified level of confidence.  The level of confidence represents the level of protection afforded to facilities whose emissions are in line with the best performers.  For example, a 99% level of confidence means that a facility whose emissions are consistent with the best performers has one chance in 100 of exceeding the floor limit.  A prediction interval for a single future observation (or an average of several test observations) is an interval that will, with a specified degree of confidence, contain the next (or the average of some other pre-specified number) of randomly selected observation(s) from a population.  In other words, the upper prediction limit estimates what the upper bound of future values will be, based upon present or past background samples taken.  The UPL consequently represents the value at which we can expect the mean of future observations for the HAP emissions to fall within a specified level of confidence, based upon the results of an independent sample from the same population.  This method accounts for the point-to-point variability in the data.  
The form of the UPL equation differs somewhat depending upon the number of data points and data distribution to which it is applied. Attachment A includes a flow diagram that summarizes the UPL approaches used. To this end, the data sets were evaluated for each HAP to ascertain whether the data were normally distributed, or fit another type of distribution (e.g., log normal distribution).  According to the Central Limit Theorem (Durrett, 1996), when a data set includes 15 or more sources, the UPL is based on the assumption that the data fit a normal distribution.  The Central Limit Theorem states that regardless of the shape of the original distribution, if the distribution has a finite mean (μ) and variance (σ²), the sampling distribution of the mean approaches a normal distribution with a mean of (μ) and a variance of σ²/N as N, the sample size, increases (Durrett, 1996). 
The wool fiberglass data sets used to calculate MACT floors varied for each pollutant. When the sample size is smaller than 15 and the distribution of the data is unknown, the Central Limit Theorem cannot be used to support the normality assumption.  Statistical test of the kurtosis and skewness are then used to evaluate the normality assumption.  The skewness statistic (S) characterizes the degree of asymmetry of a given data distribution.  Normally distributed data have an S value of 0.  An S value that is greater (less) than 0 indicates that the data are asymmetrically distributed with a right (left) tail extending towards positive (negative) values.  The standard error of the skewness statistic (SES) was also used in determining the normality of the data distribution.  The kurtosis statistic (K) characterizes the degree of peakedness or flatness of a given data distribution in comparison to a normal distribution.  Normally distributed data have a K value of 0.  A K value that is greater (less) than 0 indicates a relatively peaked (flat) distribution.  The standard error of the kurtosis statistic (SEK) was also used in determining the normality of the data distribution.
For each data set to which the UPL was applied (i.e., the separate data sets for each HAP applicable to a source), the S and K values were calculated using the reported test values.  If both kurtosis and skewness tests indicate the data is normally distributed, the UPL was calculated using the UPL pooled variance Equation 1.  

                     UPL= XT+ t(p,df)xs21n+1m  		Equation 1

where:

	XT 	=  the average (mean) of the best performing existing sources;
	t(p,df) 	=  the t statistic for a confidence level p, and df degrees of freedom;
	s[2]	=  the pooled variance;
	n	=  the total number of test runs (all sources) used in the analysis; and
	m	=  the number of (future) compliance test runs [for run-by-run data, m=3].

      Degrees of freedom calculated by:
      							    Equation 1a
      
      Mean calculated by: 
                                                                                             Equation 1b

      Pooled variance calculated by:
                                                           		    Equation 1c

If the kurtosis or skewness tests indicate the data was not normally distributed, the data were log-transformed. Once the logs of all the test runs were calculated, new kurtosis and skewness tests were performed on the log-transformed data. If both kurtosis and skewness tests indicated the log-transformed data were normally distributed, the UPL can be calculated using the Equation 2.  

                          Equation 2
where:

	μ	=  the average of the best performing existing sources;
	 	=  the z statistic for a lognormal distribution at 99 percent;
		=  the variance;
	n	=  the total number of test runs (all sources) used in the analysis; and
	m	=  the number of (future) compliance test runs [for run-by-run data m=3].				        	

      Mean is calculated by:
                                                                                                      Equation 2a
      Variance is calculated by:
                                                                                         Equation 2b

If the raw data and the log-transformed data were not normally distributed and n >=13, the UPL was calculated using the UPL pooled variance with skewness adjustment in Equation 3.  Note this adjustment cannot be used if the number of individual runs is less than 13.  

                                                                            Equation 3

where:

	XT 	=  the average of the best performing existing sources;
	t(p,df) 	=  the t statistic for a confidence level p, and df degrees of freedom;
	s[2]	=  the pooled variance;
	n	=  the total number of test runs (all sources) used in the analysis;
	m	=  the number of (future) compliance test runs [for run-by-run data m=3]; and
	Skew	=  the skewness of the test runs used in the analysis.
	
If the raw data and the log-transformed data were not normally distributed and n was less than 13 the UPL was calculated using the UPL pooled variance in Equation 1, but instead of using the Excel TINV formula to find the t statistic t(p,df), a trial and error method was used to find the t statistic that gives a 99 percent confidence level by correcting the probability values  (p) using Equation 4

Corrected Probability = P0(t) + Skewness x Pλ3(t)  -  Kurtosis x Pλ4(t) + λ3[2]xP λ3/2(t)       Equation 4

Adjustment for Below Detection Level Emissions Data
Prior to calculating the UPLs, test runs that were below detection level (BDL) were identified.  If a dataset contained any BDL values, the UPL was calculated using BDL values as reported in the test report. The calculated floor or emissions limit was then compared to three times the RDL values provided by EPA. If the three times the RDL was less than the calculated floor or emissions limit (e.g., calculated from the UPL), we conclude that measurement variability was adequately addressed and the calculated floor or emissions limit would need no adjustment. If, on the other hand, the value equal to three times the RDL were greater than the UPL, we concluded that the calculated floor or emissions limit does not account entirely for measurement variability and used the three times RDL value instead. The RDL for each pollutant was determined using data from tests of all the best performers for all of the final regulatory subcategories (i.e., pooled test data). If no BDL values were included in the pollutant data set for a particular unit type, these steps were not performed.
MACT Floor Analysis Results
The MACT floors for new and existing sources were calculated using the 99% UPL approach described above and are summarized in Table 1. When calculated MACT floors for new sources are less stringent than the MACT floor calculated for existing sources, the existing MACT floor calculated value was used for both new and existing sources. 
               Table 1  -  Summary of MACT Floor Emission Limits

                                       
                               Data Distribution
                                Mean of Sources
              Used to Calculate UPL (lb/ton) for Existing Sources
               MACT floor limits for  Existing Sources (lb/ton)
                  MACT floor limits for New Sources (lb/ton)
RS Lines
                                       
                                                                   Formaldehyde
                                    Normal
                                     0.15
                                     0.22
                                     0.087
                                                                         Phenol
                                    Normal
                                     0.15
                                     0.31
                                     0.034
                                                                       Methanol
                                    Normal
                                     0.42
                                     0.75
                                     0.61
FA Lines
                                       
                                                                   Formaldehyde
                                    Normal
                                     3.53
                                     5.55
                                     3.32
                                                                         Phenol
                                    Normal
                                     0.51
                                     1.36
                                     0.46
                                                                       Methanol
                                    Normal
                                     0.30
                                     0.50
                                     0.50

References: 
Durrett, Richard (1996). Probability:  Theory and Examples (Second edition).
Bhaumik, D.K. and Gibbons, R.D. 2004.  An Upper Prediction Limit for the Arithmetic Mean of a Lognormal Random Variable.  Technometrics, 46(2):239-248.
Johnson, Norman J.,  Modified t Tests and Confidence Intervals for Asymmetrical Populations. Journal of the American Statistical Association, Vol. 73, No. 363 (Sep., 1978), pp. 536-544.
                                 Attachment A
                                       
                                       
