Evaluation of Best Management Practices for Highway Runoff Control (2006)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

56 8.1 Methodology Options A variety of methods may be used to assess the efficiency and effectiveness of a given BMP, but it is important to under- stand the constraints and assumptions of any method before selecting it. For example, many effectiveness models measure efficiency on a storm-by-storm, pollutant-load basis, which assumes that for any given storm, the influent and effluent volumes are equal, and the outflow is directly related to the inflow. However, many systems using a permanent pool of water (wetlands and wet ponds) may not experience complete permanent pool displacement during all storm events, which invalidates storm-by-storm comparisons for removal per- centages (Strecker et al. 2001). Another factor that may influence performance measures of a given BMP, especially if the performance measure is an efficiency ratio or removal percentage, is the influent concen- tration. The efficiency ratio, which is one definition for removal fraction, R, is where the inflow and outflow EMCs are usually defined as averages of the individual events (averages of n sampled storm events, for instance). When using the efficiency ratio method, a lower influent concentration may lead to the mis- characterization of a BMP as less effective because the relative removal percentage may be far less than if a higher influent concentration was observed. Thus, as influent concentrations vary, the relative removal percentage may be a measure of the cleanliness of the influent, not the performance of the system. Especially for systems reliant on settling as their primary removal process (e.g., ponds), TSS “removal” is almost entirely a function of influent concentration combined with particle size distribution (Strecker et al. 2001). That is, TSS effluent concentrations can range widely for most ponds, Efficiency ratio Removal fraction EMCou   R = −1 t inEMC -/ 8 1( ) and the ratio of effluent to influent concentration simply decreases as the influent concentration increases. Because of its reliance upon influent concentrations, the efficiency ratio method does not account for the performance of a BMP that results in relatively constant effluent EMC levels independent of inflow conditions (e.g., media filters). Currently, the Inter- national BMP Database (www.bmpdatabase.org ) employs the efficiency ratio method to calculate the relative perfor- mance of the BMPs, but the robustness of this method is lim- ited, as just mentioned, and does not provide a process level characterization of the system performance. An abridged discussion of alternative methods for evalua- tion of BMP EMC data is included in Section 8.5. A broader discussion may be found in GeoSyntec et al. 2002, Brown 2003, and Minton 2005. 8.2 Use of EMC Most available monitoring information is in the form of EMCs, used to depict the average, flow-weighted concentra- tion of a constituent over the total length of a storm event. EMCs can be used to assess not only solids and sediment but also most other constituents (nutrients, metals, and hydro- carbons) as well. EMC is defined as the total pollutant load (mass) divided by the total runoff volume for an event of specified duration, as indicated in Equation 8-2: In Equation 8-2, C(t) and Q(t) are the time-variable con- centration and flow measured during a runoff event. M repre- sents the total pollutant mass, and V represents the total runoff volume. In practice, EMCs are determined from a laboratory analysis of a flow-weighted composite of several samples col- lected during a storm. Instantaneous concentrations during EMC C M V C(t)Q(t)dt Q(t)dt -= = = ( )∫ ∫ 8 2 C H A P T E R 8 Performance Evaluation

57 an event may vary greatly from the calculated EMC, but use of the EMC ensures that the relative mass of pollutant in a sys- tem during a storm is accurately represented (Huber 1993). EMCs are also the primary means of reporting monitoring information within the International BMP Database because of the highly variable and error-prone nature of intrastorm data collection (grab samples). The variability among EMC data is generally statistically defined by either the standard deviation or coefficient of variation (Minton 2005). 8.3 Use of the International BMP Database Web Site for Data Acquisition The primary goals of the International BMP Database project are to facilitate efficient data entry, provide useful queries of stored data, and deliver relevant performance information in a comprehensive and applicable manner through a user-friendly interface (Strecker et al. 2001). The project, which began in 1996 under a cooperative agreement between ASCE and the USEPA, now has support and funding from a broad coalition of partners including the Water Envi- ronment Research Foundation (WERF), ASCE’s Environ- mental and Water Resources Institute (EWRI), FHWA, and the American Public Works Association. Wright Water Engi- neers, Inc. and GeoSyntec Consultants maintain and operate the database and web site. The web site (www.bmpdata base.org) and its related documents are a comprehensive col- lection of reports and analyses of water quantity and quality measurements previously conducted for a variety of BMP performance studies (Quigley et al. 2002). Evaluation of the BMP data to date has indicated that BMP pollutant-removal performance is best assessed by determination of • The amount of runoff that is prevented, i.e., disposed of on-site, • The amount of runoff that is captured/bypassed by the BMP, and • The effluent quality of the treated runoff (Quigley et al. 2002). Within this chapter, summary responses to these three items are presented. Methods for analysis of quality data are not included herein. Refer to GeoSyntec et al. (2002) and Burton and Pitt (2002) for a thorough discussion of statisti- cal and other data analysis methods. Issues associated with use of the data include obtaining a large enough number of samples to achieve statistically sig- nificant results, acquisition of correct flow measurement data, and the typographical and organizational formatting prob- lems that arise when data are submitted by many contribu- tors (Quigley et al. 2002). A summary (April 2004) is presented in Table 8-1 of statis- tical analyses that are available from the International BMP Database team for some common stormwater parameters and for BMP categories that have enough data to allow for useful analysis to be conducted relative to the scope of this project. For example, the range of locations at which statisti- cal analyses have been performed for suspended solids is 11 (for detention ponds) to 21 (for retention ponds). In other words, there are enough performance data to compare differ- ent BMP types with each other using box plots and other sta- tistical criteria. Although most of these locations will not be highway sites, the comparisons should still be relevant to highway applications. Results of these comparisons are pre- sented for individual pollutants in the Pollutant Fact Sheets, included in Appendix A of the Guidelines Manual. 8.4 Search for Intra-Event Data Early in this study, a considerable effort was devoted to searching for good intra-event data for a variety of BMPs for use in characterizing fundamental unit processes at work in the devices. In addition to a complete physical description of the BMP and its catchment, an ideal data set would include influent and effluent hydrographs and polluto- graphs, stormwater treatability, and an indication of the character of the water within the device during an event. Such a data set would include several storms. No such ideal data set was found, but 11 candidates were selected from a thorough review of the International BMP Database (see Table 8-2). These 11 sites include 21,134 individual data records. Table 8-3 and Table 8-4 are expanded versions of Table 8-2, including information on the respective studies from which the data were extracted, the type of data avail- able in each study, and other relevant information such as the number of storms studied, water-quality constituents examined, a catchment description, and notes regarding data collection and sampling. All but one of the studies (Dayton Ave. biofilter) include both flow and water-quality intra-event data. With the excep- tion of the U.S. 183, Walnut Creek, and Seton Pond data, which were made available electronically by Dr. Michael Barrett at the University of Texas at Austin, all of the intra- event data were entered by hand from hard copies of the respective studies or, in the case of hydrologic data for the Moyewood Pond and Dayton Ave. biofilter, digitized from hard copy graphical displays. All hand-entered data were checked for errors, record by record. The maximum values for each field were also examined to minimize errors caused by missing decimal points (the most common mistake identi- fied). Assessment of these corrections indicates that the per- centage of identified errors in hand-entered data was less than 1% for most studies.

Parameter Name BMP Category Data Solids, Total Suspended Cu, Dissolved Cu, Total Zn, Dissolved Zn, Total Pb, Dissolved Pb, Total Cd, Dissolved Cd, Total P Total Nitrate Nitrogen, Total Nitrogen, Ammonia Total Nitrogen, Kjeldahl, Total Nitrogen, Total Biofilter Count of Inflow n 17 11 14 11 17 11 17 7 8 18 15 1 12 4 Count of Outflow n 17 11 14 11 17 11 17 7 8 18 15 1 12 4 Detention Basin Count of Inflow n 11 6 12 6 13 6 12 4 7 10 7 6 4 Count of Outflow n 11 6 12 6 13 6 12 4 7 10 7 6 4 Hydrodynamic Device Count of Inflow n 13 6 9 6 11 6 8 4 5 9 2 4 3 1 Count of Outflow n 13 6 9 6 11 6 8 4 5 9 2 4 3 1 Media Filter Count of Inflow n 18 16 18 16 18 16 18 8 9 17 16 8 15 Count of Outflow n 18 16 18 16 18 16 18 8 9 17 16 8 15 Retention Pond Count of Inflow n 21 4 13 4 17 5 16 1 10 20 4 9 12 6 Count of Outflow n 21 4 13 4 17 5 16 1 10 20 4 9 12 6 Wetland Basin Count of Inflow n 12 1 2 2 7 3 6 1 2 13 6 8 6 10 Count of Outflow n 12 1 2 2 7 3 6 1 2 13 6 8 6 10 Source: International BMP Database (http://www.bmpdatabase.org/) April 2004. Table 8-1. Number of detailed statistical analyses by common stormwater parameter for various BMP categories.

59 Generally, there were two primary issues affecting the usabil- ity of the data for the International BMP Database. First, several of the studies had sparse data or large sections of absent data for certain water-quality constituents and/or certain monitored runoff events. The proportion of records with missing water- quality and/or flow data was particularly high for the Queen Anne’s Pond study.Others with some missing data were the U.S. 183 study, the Walnut Creek study (gaps in concentrations of bacteria and some metals), and the Seton Pond study (missing concentration, flow, and/or time of sample data). For all the studies, records with no reported concentration or load, and/or no flow data were not considered for this project. Second, several studies, particularly the Queen Anne’s Pond study, included flags on reported data (such as letters next to sample IDs and/or flags on water-quality data), but no explanation of their meaning. For all studies, the following assumptions were made for commonly encountered flags: • Data such as “<10” indicate values known to be less than value shown, • Data such as “>10” indicate values known to be greater than value shown, • Data such as “–0.3” indicate non-detects, with the reported value representing the detection threshold, and • Data reported as “NA” or “—” indicate missing data. In spite of gaps in bacteriological and some metals data, three sites in Austin, Texas (U.S. Hwy. 183, Walnut Creek, and Seton Pond), and the site on Moyewood Pond in Greenville, North Carolina, had the best documented and most complete data. Hydrologic and water-quality data from these sites in Austin and Greenville were used for preliminary assessment of EMC evaluation techniques and for testing of the SWMM for application for hydrologic screening (Brown 2003). Another potential source of intra-event BMP performance information is proprietary BMP makers. However, virtually all of the data are EMCs and the quality-assessment/quality- control (QA/QC) procedures for these studies are generally not well documented. Other research investigated during the course of this study include research conducted by the Wisconsin Department of Natural Resources (Personal communication, Roger Banner- man, Wisconsin Department of Natural Resources, Madison, Wisconsin, 2003; www.dot.wisconsin.gov/library/research/ docs/quarterlyreports/0092-00-03.pdf), City of Griffin, Georgia (Keller 2002), the U.S. Geological Survey literature review of highway-related stormwater studies (Granato 2003), and the Runoff Water Quality Knowledge Base for Win- dows CD-ROM prepared by GKY Associates (2001). Results from the first two studies were not available in time to be included in this report, and none of these sources yielded any better data sources than the ones from the International BMP Database described above, although future review is war- ranted. Some Caltrans data were used later in this study for evaluation of swales, filter strips, and detention ponds. Essen- tially, evaluation of all the data is incorporated into the Pol- lutant Fact Sheets, included in Appendix A of the Guidelines Manual. A broad summary is presented in Section 8.6. Sec- tion 8.5 provides descriptions of recommended methods for evaluating the effectiveness of individual BMPs. The methods are illustrated using data from one of the three Austin, Texas, studies discussed above. Data from all three Austin sites and BMP Name Location BMP Type Hydrologic Data Water Quality Data Seattle METRO Retention Pond Bellevue, WA Retention pond Yes Yes Whispering Heights Residential Pond Bellevue, WA Detention basin Yes Yes Moyewood Pond Greenville, NC Detention basin Yes* Yes Demonstration Urban Stormwater Treatment (DUST) Marsh System A Freemont, CA Wetland channel Yes Yes DUST Marsh System C Freemont, CA Wetland channel Yes Yes Barton Creek Square Shopping Center Pond Austin, TX Detention basin Yes Yes U.S. 183 at MoPac Expressway, Grass Filter Strip Austin, TX Biofilter Yes Yes Walnut Creek Vegetative Buffer Strip Austin, TX Biofilter Yes Yes Seton Pond Sedimentation Facility Austin, TX Detention basin Yes Yes Queen Anne’s Pond Centerville, MD Wetland channel Yes Yes Dayton Ave. Biofilter—Grass Swale Seattle, WA Biofilter Yes* No *Hydrograph data digitized from graphical displays Table 8-2. BMPs with intra-event data as of July 2003.

DUST Marsh System C WC Urban Stormwater Treatment at Coyote Hills Marsh ABAG 1986 7 pH, EC, TDS, TSS, BOD, oil/grease, NH4, NO3, KN, Ortho-P, TP, Cd, Cr, Cu, Pb, Mn, Ni, Zn general urban (Fremont, CA) Yes Yes Yes No Barton Creek Square Shopping Center Pond DB Effects of Runoff Controls on the Quantity and Quality of Urban Runoff at Two Locations in Austin, Texas Welborn and Veenhis 1987 6 COD, BOD, fecal coliform strep, DS (180 degC), DS (105 degC), VDS, TN, NO2+NO3, NH4, POC, Cd, Fe, Pb, Zn shopping center Yes Yes No No U.S. 183 at MoPac Expressway —Grass Filter Strip BF Use of Vegetative Controls for Treatment of Highway Runoff Walsh et al. 1997 15 TSS,turbid, TSS,turbid, TSS,turbid, fecal colif, strep, ecoli, COD, TOC, NO3, TKN, TP, Zn, Pb, Fe, Cu highway Yes Yes No No Walnut Creek Veg. Buffer Strip BF Use of Vegetative Controls for Treatment of Highway Runoff Walsh et al. 1997 23 fecal colif, strep, ecoli, COD, TOC, NO3, TKN, TP, Zn, Pb, Fe, Cu highway Yes Yes No No TSS BMP Name BMP Type Study Name Author(s) Approx. # of Storms w/Sufficient Data Water Quality Constituents Catchment Description Hydro- logic Data WQ Data Particle Size Distribution? Settling Velocity? Seattle METRO Retention Pond RP Operation of Detention Facilities for Urban Stormwater Quality Enhancement Dally et al. 1983 6 grease and oil, grease and oil, TSS, TP, TCd, SolCd, TPb, SolPb, TZn, SolZn, bus parking lot/ maintenance Yes Yes No No Whispering Heights Residential Pond DB Operation of Detention Facilities for Urban Stormwater Quality Enhancement Dally et al. 1983 4 TSS subdivision Yes Yes No No Moyewood Pond DB An Evaluation of Pollutant Removal by a Demonstration Urban Stormwater Detention Pond Stanley 1994 8 TSS, VSOL, FSOL, NH4, NO3, PO4, TKN, TDP, PN, PP, DOC, POC, Cr, Cd, Ni, Pb, Cu, Zn, fecal coliform BOD, COD residential and commercial Yes (see Comme nts) Yes No Yes Demonstra- tion Urban Stormwater Treatment (DUST) Marsh System A WC Urban Stormwater Treatment at Coyote Hills Marsh ABAG 1986 7 pH, EC, TDS, TSS, BOD, oil/grease, NH4, NO3, TKN, Ortho- P, TP, Cd, Cr, Cu, Pb, Mn, Ni, Zn general urban (Fremont, CA) Yes Yes Yes No Seton Pond Sedimenta- tion Facility DB The Effectiveness of Permanent Highway Runoff Controls: Sedimentation/Filtration Systems Keblin et al. 1997 9 COD, TOC, NO3, TKN, TP, Zn, Fe highway Yes Yes No No Queen Anne's Pond WC The Use of Artificial Wetlands in Treating Stormwater Runoff Athanas and Stevenson 1991 14 NO3, NO2, TN, TDN, PO4, TP, TDP, TSS, ON, OP, PN, PP, NO3+NO2, NH4 high school Yes Yes No No Dayton Ave. Biofilter— Grass Swale BF Dayton Avenue Swale Biofiltration Study Goldberg et al. 1993 7 No pollutographs urban (Seattle, WA) Yes (see Table 8-4) No No No Note: RP = retention pond, DB = extended detention basin, WC = wetland channel, BF = biofilter, T = total, Sol = soluble, EC = electrical conductivity, VSOL =volatile solids, FSOL = fixed solids, TKN = total Kjeldahl nitrogen, TDP = total dissolved phosphorus, PN = particulate nitrogen, PP = particulate phosphorus, DOC = dissolved organic carbon, POC = particulate organic carbon, DS = dissolved solids, VDS = volatile dissolved solids, TN = total nitrogen, TDN = total dissolved nitrogen, ON = organic nitrogen, OP = organic phosphorus, PN = particulate nitrogen, and PP = particulate phosphorus. Table 8-3. Extended summary table of candidate intra-event sites.

61 Barton Creek Square Shopping Center Pond WQ and discharge data collected immediately up- stream and downstream of pond. Manning X System Level Transmitter and Recorder was used at the inflow and outflow stations to measure stage. Manning S- 4050 automatic water samplers were used to collect WQ samples. The sampler intake was located near the bottom of the channel and was activated when the stage rose to a predetermined level (not specified). Two one-liter bottles were sampled at every interval. Data difficult to interpret. No clear methodology for when multiple intra- event samples taken. Extracted data limited to those events where >3 intra- event data are clearly evident. U.S. 183 at MoPac Expressway, Grass Filter Strip Isco 3700 samplers and Isco 3230 bubbler flow meters used to measure water quality and flow at the inlet and outlet of the filter strip. Data provided electronically by Michael Barrett. Inflows and outflows have different number of storms. Several records missing flow and/or water quality data. Walnut Creek Vegetative Buffer Strip Isco 3700 samplers and Isco 3230 bubbler flow meters used to measure water quality and flow at the inlet and outlet of the filter strip. Data provided electronically by Michael Barrett. Inflows and outflows have different number of storms. Several records missing flow and/or water quality data. Seton Pond Sedimentation Facility Isco 3700 sampler drew inflow samples at 15, 45, 75, 135, and 195 min, after flow depth in influent channel reached 2.5 cm, and every 60 min thereafter, if necessary. Isco 3700 sampler drew outflow samples at 5, 120, 240, 600 and 960 min after activation. Isco 3230 bubbler flow meters measured flow at inlet and outlet. Data provided electronically by Michael Barrett. Inflows and outflows have different number of storms. Several records missing flow and/or water quality data. Queen Anne's Pond No information available. Irregular data. Data from both inlets not provided for many storms. No annotation explaining flags marking many records, unclear whether these are dupes, problematic samples, or other. Only flow volumes provided for majority of storms (no time data), with no clear way of translating into hydrograph. Dayton Biofilter— Grass Swale Flow measured using an automatic flow meter connected to a pressure transducer, installed at the inlet and outlet (H-flumes). Hydrologic data in graphical form only. Points digitized and flow volumes for each sample interpolated. No "start of event" time given in study, so assumed that lower scale of x-axis represented beginning of event. One of the eight monitored storms not used due to known obstruction in flow path, affecting flow data. BMP Name Sampling Method Comments Seattle METRO Retention Pond Outflow measured in V-notch flume. Pond height measured using nitrogen gas pressure differential. Stevens A-35 water level recorder and stilling well were installed to record flume stage. Manning 4040 discrete water quality sampler was installed to sample discharge from the flume. The samplers are capable of taking 24 one-liter samples at carrying time increments. Good data, some printing errors identified. Study text details lab calibration of special stage-discharge relationship for flume used to measure flow. Whispering Heights Residential Pond Stevens A-35 float type water level recorder in 12” stilling well used to measure the water level in the pond. 6” orifice at outlet controlled discharge. Manning 4040 discrete automatic samplers with adjusted one-liter intake capacities were used. Outflow samples were collected directly upstream of outlet. Good data, some printing errors identified. Study text details lab calibration of special stage-discharge relationship used to measure flow based on pond surface elevation. Moyewood Pond WQ sampled automatically (ISCO Model 2700). Inflow at 15-min intervals for first 2 hrs and every hour thereafter. Outflow every hour. Every other WQ sample discharged when delta TSS was "small" in order to reduce lab costs. Hydrologic data in graphical form only. Points digitized and flow volumes for each sample interpolated. Two pages of data missing (end of storm 7, beginning of storm 8) and pages missing from summary of analysis methodology in text. Water quality sampling period does not always exactly overlap with available flow data; as a result, several water quality records have no corresponding flow data. Demonstration Urban Stormwater Treatment (DUST) Marsh System A Flow rate computed from stage-discharge relationship. Water quality measurements from grab samples. DUST Marsh A and C in series. A is 5-acre lagoon system. Time of samples provided as "sampling periods," separately from tabulated data. It is assumed that each sample was taken at the end of the corresponding sampling period. DUST Marsh System C Flow rate computed from stage-discharge relationship. Water quality measurements from grab samples. DUST Marsh A and C in series. C is wetland channel. Treats flows from System A and System B. No inlet data for System B, only outlet data. Time of samples provided as "sampling periods," separately from tabulated data. It is assumed that each sample was taken at the end of the corresponding sampling period. Table 8-4. Sampling method and additional comments on candidate intra-event sites.

62 the Greenville site were evaluated by Brown (2003), whose study may be consulted for more detail. 8.5 Evaluation of Quality Performance for Individual BMPs 8.5.1 EMC Data for U.S. 183 Filter Strip As described in the previous section, University of Texas researchers (Walsh et al. 1997) collected performance data for a major highway median filter strip in northwestern Austin, Texas (see Table 8-5 and Figure 8-1). Although most of the data were obtained from the International BMP Data- base, considerable help in obtaining and interpreting the data was provided by Professor Michael Barrett of the Uni- versity of Texas. Data for three parameters—TSS, nitrate, and total zinc—will be used to illustrate some EMC evaluation procedures, although several other constituents were sam- pled as well (see Table 8-3). These three parameters were chosen because they were available for multiple studies eval- uated during the project. The basic data used for the com- parisons that follow are shown in Table 8-6. There are at least 19 EMC data points for each of the three parameters, which is more data than is usually available for BMP evaluation studies. Note, however, that flows were not sampled for all events for which quality was sampled and that more effluent samples were collected than influent samples. The results of simple statistical analysis of the data are also shown in Table 8-6. The total zinc data illustrate a detection limit issue, with seven effluent EMCs at 0.002 mg/L. Methods of dealing with data at the detection limit are given by Burton and Pitt (2002). For instance, because of the huge data-handling requirements within the International BMP Database, all records flagged as nondetects are assigned one-half the reported detection limit, although Helsel and Hirsch (1992) advocate more robust methods based on the characteristics of observed data greater than the detection limit. However, following the methods of Walsh et al. (1997), detection limit data points herein are kept at detection limit values in order to be conservative (higher EMCs than might otherwise be occurring) for effluent BMP quality. Simple removal efficiencies (see Equation 8-1) are also shown in Table 8-6, based on average effluent and influent EMCs. The efficiency ratio weights all EMCs equally, regard- less of the magnitude of the storm, and therefore will yield inconsistencies if the EMC varies significantly with the storm volume or if the pollutant loads are not necessarily propor- tional to the storm volume. For the three tabulated pollutants, however, removal does occur. 8.5.2 Scatter Plot Methods for analysis of EMC data have been evaluated in detail by GeoSyntec Consultants et al. (2002), Huber et al. (2006), and Minton (2005). Prior to an evaluation involving a “removal” estimation or formal statistical procedure, or prior to the more sophisticated effluent probability method used in this study, a simple plot of storm event effluent EMC versus influent EMC should be conducted for a qualitative estimate of effectiveness. Hypothetical data are shown in Figure 8-2 to illustrate possible relationships. If the data plot at less than a 45° line (effluent = influent), some removal is occurring. If removal occurs and the line is approximately lin- ear, then where R = removal fraction. If the relationship is nonlinear, particularly, if apparent removal increases with higher influ- ent concentration (as illustrated in Figure 8-2), then it can be inferred that removal is a function of influent concentration. However, if effluent EMCs are relatively constant (the square data points in Figure 8-2) and not a function of influent EMCs, then “percent removal” is a poor way to characterize the BMP performance, and effluent quality itself is the key characteristic. This property (EMC distribution) may in turn EMC R EMC -out in= −( )⋅ ( )1 8 3 Characteristic U.S. 183 Centerline length (m) 356 Width of entire median (m) 14.9-19.5 Filter strip treatment length (m) 7.5-8.8 Average median side slope (%) 12.10% Average centerline slope (%) 0.73% Drainage area (m2) 13,000 Average daily traffic 111,000 Impervious drainage area (%) 52% Impervious roadway area (%) 100% Source: Brown 2003, using data from Walsh et al. (1997). Table 8-5. Characteristics of the U.S. 183 filter strip. Source: Walsh et al. 1997. Figure 8-1. Vegetated filter strip at U.S. 183 site.

63 y = 0.7148x 0 20 40 60 80 100 120 140 160 0 50 100 150 200 EMCin, mg/L EM Co ut , m g/ L 1) EMCout = EMCin*frac 45-deg line, effluent = influent 2) EMCout = EMCin^exponent 3) No relationship PowerFunction Linear (1) EMCout = EMCin*frac) y = x^0.85 Figure 8-2. Scatter plot of hypothetical data to illustrate possible relationships between effluent and influent EMCs. Storm No. Date Inflow Volume, L Outflow Volume, L TSS in mg/L TSS out mg/L NO3* in mg/L NO3 out mg/L T-Zn** in mg/L T-Zn out mg/L 9 5/27/1996 --*** 60,390 -- 4 -- 0.65 -- 0.011 10 4/5/1996 -- 74,350 -- 18 -- -- -- 0.003 11 4/22/1996 -- 42,120 -- 5 -- 0.8 -- 0.002 12 5/27/1996 117,310 148,880 127 56 1.21 0.54 0.294 0.002 0.279 0.022 13 5/30/1996 176,250 254,830 7 56 5.21 4.61 0.115 15 6/22/1996 6,800 62,330 247 38 3.29 2.71 0.459 0.002 16 6/25/1996 18,480 115,950 117 50 5.66 3.71 0.285 0.003 18 8/11/1996 -- 170,400 -- 58 -- 0.64 -- 0.002 19 8/22/1996 5,940 35,410 31 3 2.66 0.31 0.002 0.002 0.044 20 8/23/1996 15,260 149,470 151,690 161,420 17 5 0.8 0.2 0.03 0.002 21 8/29/1996 3,680 22 59 1.12 1.4 0.146 22 9/18/1996 32,320 103,090 135 7 2.25 1.32 0.123 0.002 23 10/17/1996 18,400 -- 64 -- 1.15 -- 1.04 -- 24 10/27/1996 6,320 -- 312 -- 0.47 -- 1.099 -- 25 11/7/1996 30,600 222,390 81 14 0.53 0.2 0.126 0.025 26 11/24/1996 25,660 294,980 40 6 0.41 0.19 0.022 0.026 28 12/15/1996 55,180 475,340 98 7 0.55 0.25 0.093 0.022 29 2/7/1997 -- -- 7 -- 1.12 -- 30 2/12/1997 10,110 257,940 133 5 0.46 0.78 0.23 0.027 31 3/11/1997 -- 122,090 -- 17 -- 0.17 -- 0.11 32 3/25/1997 20,490 175,860 328 6 0.43 0.41 0.44 0.07 33 4/2/1997 12 ,360 65,380 522 6 1.63 0.68 0.69 0.05 34 4/25/1997 3,830 426,570 146 4 2.47 0.32 0.35 0.07 35 5/9/1997 152,400 462,050 389 21 0.91 0.42 0.48 0.05 36 5/27/1997 87,389 301,330 159 19 0.94 0.48 0.41 0.07 n 19 23 19 23 19 22 19 23 average 42,041 188,446 157 20.5 1.692 0.996 0.347 0.032 std. dev. 52,456 130,775 141 20.7 1.568 1.181 0.313 0.035 CV**** 1.248 0.694 0.903 1.009 0.927 1.186 0.901 1.091 median 18,480 151,690 127 7 1.12 0.59 0.285 0.022 R (Eqn. 8- 1) -- -- -- 0.87 -- 0.41 -- 0.91 * NO3 = nitrate ** T-Zn = total zinc *** -- = no available data. ****CV = coefficient of variation = standard deviation / average. Source: Walsh et al. 1997. Table 8-6. Water-quality data (EMCs) for the U.S. 183 filter strip.

64 be characterized by a frequency distribution, as exemplified by the effluent probability method (EPM) (see Section 8.5.3). In any event, scatter plots are useful for such qualitative evalua- tion and are shown for the U.S. 183 data in Figures 8-3, 8-4, and 8-5. In the scatter plots of TSS and total zinc for U.S. 183, the effluent concentrations are essentially flat and seemingly not influenced by the influent concentration (see Figures 8-3 and 8-5, respectively). No functional removal relationship is apparent; what is apparent is that effluent EMCs are much lower than influent EMCs The frequency distributions pro- vided by the EPM are an excellent means to provide such a characterization of effluent EMCs. With respect to nitrate (see Figure 8-4), removal is less, which is to be expected for a dissolved constituent. A trend line, forced through zero, has a relatively high R2 value, but the statistical significance cannot be evaluated (because of the forced zero intercept). A trend line not forced through zero is y = 0.202 + 0.719x, for which R2 = 0.78 and for which p = 3  106 (highly significant). However, the intercept is not signif- icant (p = 0.41). Hence, one definition of removal fraction for nitrate would be 1 – slope, in the range of 0.28 – 0.35 (at this site only; nitrate is poorly removed by most BMPs). These val- ues are somewhat lower than the efficiency ratio removal fraction computation of 0.41 (see Table 8-6). Although an EMC reduction does occur for the majority of BMPs and constituents (see the Guidelines Manual, Appendix A,“Pollutant Fact Sheets”), the relationship can often only be characterized by a significant difference between EMC medi- ans or through the use of other nonparametric statistical tests. These tests will also be illustrated below. 8.5.3 EPM The EPM is straightforward and provides a clear, but qual- itative picture of BMP effectiveness. The EPM consists of a lognormal probability plot (although any distribution could be used for which probability paper exists, including normal) of EMC versus either probability of occurrence or percent exceedance (equivalent to the cumulative distribution func- tion). Probability plots are among the most useful pieces of information that can result from a BMP evaluation study (Burton and Pitt 2002). The authors of Urban Stormwater BMP Performance Monitoring: A Guidance Manual for Meet- ing the National Stormwater BMP Database Requirements (GeoSyntec et al. 2002) strongly recommend that the stormwater industry accept probability plots as a standard “rating curve” for BMP evaluation studies as they provide a visual representation of the frequency distribution of both influent and effluent quality. Lognormal plots are ordinarily used because the lognormal distribution has been found to be a good fit for most stormwa- ter EMC data (USEPA 1983; Driscoll 1986; Smullen and Cave 2002; George 2004), although the normal distribution has been found to be a better fit for pond effluent data in some studies (Van Buren et al. 1997). One advantage of normality, either of the logs or of the untransformed data, is that parametric 0 20 40 60 80 100 120 140 160 180 200 0 100 200 300 400 500 600 Influent TSS, mg/L Ef flu en t T SS , m g/ L Effluent = Influent y = 0.6559x R2 = 0.7658 0 1 2 3 4 5 6 0 2 4 6 Influent NO3-N, mg/L Ef flu en t N O 3- N , m g/ L Effluent = Influent Note: Trend line intercept is forced through zero. Figure 8-4. Effluent versus influent nitrate (NO3-N) for U.S. 183 filter strip. Figure 8-3. Effluent versus influent TSS EMCs for U.S. 183 filter strip. 0 0.05 0.1 0.15 0.2 0 0.2 0.4 0.6 0.8 Influent Total Zn, mg/L Ef flu en t T ot al Z n, m g/ L Effluent = Influent Figure 8-5. Effluent versus influent total zinc for U.S. 183 filter strip.

65 statistical tests can be applied, such as the t-test, chi-square test, and analysis of variance. Statistical tests used to compare data sets typically require normality in the data sets, and some also require the data sets to have equal variances. Lognormal (and normal) probability plots can also be used for qualitative guid- ance because the occurrence of two curves with the same slope indicates the same variance of the data. The most basic test for normality is whether or not the data plot as a straight line on normal probability paper (or versus equivalent values of the standard normal variate, z) (Burton and Pitt 2002; Bedient and Huber 2002). Tests for normality itself include tests directly related to probability plots, such as the probability plot correlation coefficient (PPCC) (Vogel 1986) and the Shapiro-Wilk test (Helsel and Hirsch 1992). Tests not related to probability plots include the Kolmogorov-Smirnov test and the chi-square test (Benjamin and Cornell 1970; Helsel and Hirsch 1992). How- ever, the latter two (Kolomogorov-Smirnov and chi-square) are less powerful in a statistical sense than tests that use prob- ability plots; moreover, the plots themselves yield great qual- itative information, as discussed below. It is interesting to note that if ranked data are plotted against standard normal variates (“z-values”) obtained from the inverse of the plot- ting position probability (Bedient and Huber 2002) using MS Excel, the linear fit of data (untransformed or loga- rithms) obtained using MS Excel’s “trend line” option pro- vides the required PPCC. The PPCC can then be tested for statistical significance (Vogel 1986; Helsel and Hirsch 1992). The critical values of the correlation coefficient (found in Vogel 1986 and Helsel and Hirsch 1992) account for the inherent correlation between two ranked data sets (order statistics). If normality, and, in some cases, equal variance, is ensured among the respective data sets, parametric tests can be employed to test the difference between the means (and medians if normally distributed) of the data sets. Parametric and nonparametric statistical tests should be conducted after the probability plots are generated to indicate if perceived differences in influent and effluent mean EMCs are statistically significant (the level of significance should be provided rather than just noting whether the result was sig- nificant, e.g., a 95% significance level). Helsel and Hirsch (1992) provide an excellent primer on parametric and non- parametric methods with applications to water resources and water quality. Many parametric and nonparametric tests are included in standard statistical software. Limited quantitative assumptions can be made simply on the basis of the effluent probability plots themselves. Unlike the X-Y scatter plots in Figures 8-2 through 8-5, probability plots arrange the data on the basis of ranked quantiles, not by event. Influent and effluent values for a given quantile (cumu- lative frequency) are assumed to be temporally independent (values associated with the same storm event would be sheer coincidence). Even if the distribution of effluent values lies wholly underneath the influent distribution, there is no guar- antee that effluent concentrations were less than influent con- centrations for every sampled event. However, the range of both influent and effluent quality can be determined on the basis of the concentration range between given percentile values. In addition, the normality and equal variance among the data sets can be qualitatively observed, although any inferences about variance must be confirmed through quantitative statistical testing. Finally, when influent and effluent EMC medians are separated by less than the standard deviation of either the influent or efflu- ent EMC data, this is a qualitative indication of minimal removal. Examples of effluent probability plots are shown for the U.S. 183 data in Figures 8-6 through 8-8. The abscissa in each plot is the standard normal variate, z, or the inverse of the cumulative distribution function, F(z), of the N(0,1) distribution (normal, with mean = 0 and vari- ance = 1), corresponding to Tabulated values of either F(z) or z1(F) may be found in any statistics book and may also be obtained from the Excel functions NORMDIST() or NORMINV(), respectively. Quartiles for 25%, 50% (median), and 75% correspond to z = -0.674, 0, and +0.674, respectively, and are shown on the three EPM plots. Regression lines (Excel trend lines) are also shown for each plot, corresponding to in which it is clear that the average of the logarithms of the EMCs (ln EMC) is the intercept, and the standard deviation of the logarithms of the EMCs (Sln) is the slope. Because Excel can only provide an exponential fit when using a vertical log scale, the exponential equation is given on the plots. Hence, the average of the logarithms is the log of the coefficient, and the standard deviation of the logarithms is the number in the exponent. For example, for TSS effluent, ln EMC = ln(12.452) = 2.52, and Sln = 0.9935. As a matter of interest, the least squares regression trend lines for these data are almost indis- tinguishable from lines that would show the method of moments fit for the lognormal distributions. Hence, the latter fits are not shown. One other aspect of the EPM method is that all influent and effluent data points are included in the analysis (all EMC values tabulated in Table 8-6) whereas only the lesser number of matched pairs (both values sampled for the ln EMC ln EMC S z -ln= + ⋅ ( )8 5 F z e du - u z 2 ( ) = ( ) − −∞ ∫ 2 2 8 4 π

66 y = 12.452e0.9935x R2 = 0.8969 y = 96.658e1.1569x R2 = 0.9545 1 10 100 1000 -2.00 -1.75 -1.50 -1.25 -1.00 -0.75 -0.50 -0.25 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 Standard Normal Variate (z) TS S EM C, m g/ L Influent Effluent 25% 50% 75% Figure 8-6. EPM plots for TSS for U.S. 183 filter strip, with quartile locations also shown. y = 0.6169e0.9531x R2 = 0.9522 y = 1.191e0.8475x R2 = 0.9457 0.01 0.1 1 10 -2.00 -1.75 -1.50 -1.25 -1.00 -0.75 -0.50 -0.25 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 Standard Normal Variate (z) N O 3- N E M C, m g/ L 25% 50% 75% Influent Effluent Figure 8-7. EPM plots for nitrate for U.S. 183 filter strip, with quartile locations also shown. y = 0.1881e1.4316x R2 = 0.8388 y = 0.0131e1.4905x R2 = 0.8588 0.001 0.01 0.1 1 10 -2 -1.8 -1.5 -1.3 -1 -0.8 -0.5 -0.3 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 Standard Normal Variate (z) Zi nc E M C, m g/ L 25% 50% 75% Influent Effluent Figure 8-8. EPM plots for total zinc for U.S. 183 filter strip, with quartile locations also shown.

67 same event) can be shown on scatter diagrams. This once again emphasizes the point that data values shown for the same quantiles do not necessarily (and are not even likely to) correspond to the same storm event. Examination of the EPM plots for TSS and total zinc (see Figures 8-6 and 8-8 and) suggests that the slopes of the influent and effluent lines are very similar, indicating simi- lar variances (of the logs) for the two parameters. Moreover, the separation of the influent and effluent frequency distri- butions is very clear, the significance of which will be tested subsequently. Interestingly, as will be shown subsequently as well, only the lognormal fit of the effluent TSS EMC distri- bution is statistically significant, that is, the logs of the influ- ent TSS EMCs fail two normality tests. The implication is that a parametric t-test may not be used to compare means, leading to use of the nonparametric Kruskal-Wallis test later. Overall, the U.S.183 filter strip shows obvious removal of TSS and total zinc, with significance tests to follow. A final element of the EPM plot for total zinc (see Figure 8-8) is the evidence of detection limit data, with seven values at 0.002 mg/L forming a flat line at the lower left part of the effluent EMC distribution. As mentioned previously, in real- ity, values may be less than 0.002 mg/L; hence, the plot con- servatively overestimates the magnitudes of effluent EMCs. The EPM for nitrate (see Figure 8-7) also shows a tendency toward similar variance of the logs, but with less separation between influent and effluent EMCs than is the case with TSS and total zinc. Both the influent and effluent log (EMC) values fail normality tests; thus, only a nonparametric com- parison may be used to test the significance of the separation of the influent and effluent EMCs. Finally, for illustrative purposes, EPM plots for the hypo- thetical data of Figure 8-2 are shown in Figure 8-9. Although the figure illustrates separation of data sets as before, there is almost no way of discerning the possibility of a functional relationship from the EPM. Only the scatter plot shows the linear (series 1), power function (series 2), and lack of rela- tionship (series 3) qualitatively evident in Figure 8-2. The qualitative inferences from the EPM may also be obtained from other descriptive statistics, such as box plots (see Section 8.5.4), as well as obtained quantitatively through the parametric t-test and nonparametric comparisons of medians. However, the EPM has the advantage of illustrating the lognormal (or other distribution) fit of the data, rather than simply certain quantiles, as with box plots. The primary problem with the EPM is that certain quantitative assump- tions, such as removal and performance at or around a cer- tain concentration value, cannot be made unless data points are entered as matched pairs (e.g., as in scatter plots of efflu- ent versus influent EMC). With the Caltrans BMP study and assessment (Caltrans 2003), this discrepancy was called out and discussed, indicating that interpretation of these plots should be performed in conjunction with related analyses, such as scatter plots. Another concern raised with the EPM is its ability to provide sufficient information regarding BMP selection. In areas requiring a set removal percentage, use of the EPM may not adequately portray whether a BMP is capa- ble of meeting that performance standard (Caltrans 2003). 8.5.4 Box and Whisker Plots Most statistical software will provide what is known as a box plot or box and whisker plot, representing quantiles, extremes, and confidence limits of the data. So-called notched box and whisker plots are used in Appendix A of the Guide- lines Manual and will be illustrated here for the U.S. 183 data. An explanation (taken from the same appendix) is shown in conjunction with Figure 8-10. 10 100 1000 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 Standard Normal Variate (z) EM C, m g/ L EMCin EMCout-1 EMCout-2 EMCout-3 Figure 8-9. EPM analysis of hypothetical data shown in Figure 8-2.

68 The notches encompass the 95% confidence interval of the median and provide a nonparametric means of assess- ing the difference between the centers of multiple distribu- tions. A logarithmic scale was determined to be best suited for plotting most data. The log-scale box plots were created utilizing the following method to calculate the upper and lower confidence levels: 1. The natural logs of the EMC are sorted in ascending order. 2. The upper and lower quartiles (i.e., the 75th and 25th per- centiles) are calculated, following Tukey (1977). 3. The confidence interval of the median is calculated based on the upper and lower quartiles, following McGill et al. (1978). 4. The inner quartile range (IQR) is defined as the difference between the upper (third or 75%) quartile (also, approxi- mately, the upper hinge) and the lower (first, or 25%) quartile (also, approximately, the lower hinge). The lower inner fence is the lower quartile minus 1.5  IQR, and the upper inner fence is the upper quartile plus 1.5  IQR. 5. By taking the exponent (value = elog), the upper and lower confidence levels are translated back to arithmetic space. These values are used to delineate the upper and lower bounds of the notch on the box plots. Many useful explanations of such arcane statistics may be found on the Web; one of many sites with explanations of box plot ingredients is available at: http://www.xycoon.com/ index.htm. For the distributions of averaged EMCs by BMP category and the distributions of individual EMCs by BMP category, the arithmetic values of the upper confidence level and lower confidence level of the median are provided in the table that accompanies each summary. All the box plots shown in Appendix A of the Guidelines Manual were prepared with SYSTAT software. Notched box and whisker plots for the three constituents of the U.S. 183 site are shown in Figures 8-11 through 8-13. The plots summarize inferences that may be made from the fre- quency distribution plots of the EPM. For instance, for influent Note: CL = confidence level. 3rdQuartile 1stQuartile Median Lower 95% CL Upper 95% CL Upper Inner Fence Lower Inner Fence Outside Value Figure 8-10. Box plot definitions. INFLUENT EFFLUENT 1 10 100 1000 TS S, m g/ L 0.1 1.0 10.0 N IT R AT E, m g/ L INFLUENT EFFLUENT Figure 8-11. Box plot comparison of influent and effluent EMCs for U.S. 183 TSS data. Figure 8-12. Box plot comparison of influent and effluent EMCs for U.S. 183 nitrate data. TSS data, the lower quartile of about 50 mg/L and the upper quartile of about 200 mg/L may also be read off the EPM plot (see Figure 8-6). The lower confidence limit on the median for effluent TSS is less than the lower quartile, resulting in the rever- sal of the lines at the bottom of the box plot. Only the influent total zinc data have any data points (one) less than the lower inner fence (evidently a detection limit value). Overlapping confidence intervals are an indication that medians are not significantly different. From the three box plot figures, it is to be expected that TSS and total zinc medi- ans are significantly different, whereas it is not clear about nitrate medians. The performance of definitive tests is dis- cussed in the next section. Because of the confidence interval presentation, it is generally easier to determine the signifi- cance of differences in medians from box plots than from the EPM plots, and the box plots nicely summarize the range of

69 data. Hence, box plots are used to concisely compare per- formance among BMPs in terms of effluent EMCs in the Guidelines Manual, Appendix A. On the other hand, the EPM plots show all data points and also illustrate the normality, or lack of it, of the data. EPM plots also give a better visual sense of the variance of the data. (When plotted with frequency on the abscissa and magnitude on the ordinate, a steeper slope means greater variance.) 8.5.5 Tests of Normality and Equality of Variance One way to evaluate BMP performance is to determine whether there is a statistically significant difference between the means (or medians) of the data sets (e.g., between influent and effluent EMCs within this chapter or among different BMPs in Appendix A of the Guidelines Manual), beyond the qualitative evaluations made from the EPM plots and box plots. Selection of the appropriate test to assess these differ- ences depends on the normality of the influent and effluent data sets, and, in some cases, on the equal variance of the two data sets. George (2004) summarizes several methods for eval- uation of normality of the data (or their log transforms), two of which, the probability plot correlation coefficient (PPCC) method and the Shapiro-Wilk test, will be demonstrated in this section. In addition, three different variance tests (the Lev- ene test, the Cochran test, and the Bartlett test) will be run on the individual log-transformed influent and effluent data sets for each constituent to test for equal variance. Based on the R-values obtained for the least squares trend line, the PPCC test was conducted to test for normality of the log-transformed influent/effluent data for U.S.183 (Vogel 1986; Helsel and Hirsch 1992). The correlation coefficient is simply the square root of the R2 value determined from the least squares trend line shown on the EPM plots (see Figures 8-6 through 8-8). The null hypothesis (H0) is that the data are normally distributed, and failure to reject the null hypothesis means that the data set may be assumed to be normal. How- ever, rejecting the null hypothesis does not prove normality especially for small sample sizes (Helsel and Hirsch 1992). A Type I error (probability of accepting a false hypothesis) sig- nificance level, alpha (α, or often “p”) of 0.05, is used for the normality assessment (Helsel and Hirsch 1992). Based on the sample size and alpha level, critical correlation coefficients can be obtained from tables provided by either Vogel (1986) or Helsel and Hirsch (1992) and compared to calculated cor- relation coefficients to determine whether or not to reject the null hypothesis. The PPCC test for normality maintains an advantage over other normality tests (Kolomogorov-Smirnov test and chi-square test) in that the probability plots may help to illustrate the results, and the test may be conducted over a continuous scale (Vogel 1986; Helsel and Hirsch 1992). George (2004) demonstrates the application of several meth- ods for evaluation of the frequency distribution of EMCs. As sample size increases, it is more difficult to show, based on the correlation coefficient, that the normal distribution cannot be rejected; this is because the critical correlation coef- ficient increases while the actual correlation coefficient may or may not increase with sample size. For instance, critical R values at the 95% level (α = 0.05) range from 0.879 (n = 3) to 0.939 (n = 15) and higher (for higher sample size n). There- fore, with small data sets (n < 25), departures from normal- ity must be large in order to reject the null hypothesis that the data are normally distributed (Helsel and Hirsch 1992). Ranked data are inherently correlated, i.e., the ranked EMCs increase monotonically with the increasing standard normal variates (a function of increasing cumulative frequencies). The correlation embedded in this comparison is included in the PPCC significance test. The numerically derived critical values for the test reflect expected “spurious correlation.” In order to compare available normality tests, results of the PPCC test may be compared with the results of the commonly encountered Shapiro-Wilk normality test, available in the StatGraphics statistical computer program. The Shapiro-Wilk test also uses probability plots to determine correlation coef- ficients, which describe the regression between the actual data and their normal variates or z values (Helsel and Hirsch 1992). The principal difference between the Shapiro-Wilk and PPCC tests is that the former compares values of R2 to critical values whereas the latter compares R-values. The StatGraphics pro- gram was used to compute Type I–error p-values for the log- transformed data for each constituent at each site, and the respective p-values were compared with the predetermined alpha value of 0.05. As in the PPCC test, in the Shapiro-Wilk test, if the p-value exceeds 0.05, then the null hypothesis that the data are normally distributed cannot be rejected; nonethe- less, being unable to reject the hypothesis that the data are nor- mally distributed does not ensure normality. INFLUENT EFFLUENT 0.001 0.010 0.100 1.000 ZI N C, m g/ L Figure 8-13. Box plot comparison of influent and effluent EMCs for U.S. 183 total zinc data.

70 As shown in Table 8-7, the hypothesis that the data are nor- mally distributed is confirmed for nitrate influent and effluent EMCs and for TSS influent EMCs. In order to use parametric tests to assess the difference between the means of two samples (influent and effluent), both the influent and effluent data sets must be normal; here, this is true only for nitrate. The Shapiro- Wilk results coincide with the PPCC results for these data, which is not necessarily the case with other data.When the two tests disagree, professional judgment (or other statistical tests) may be used to assess normality. Although it is not always required, equal variance between two data sets is sometimes necessary for a number of para- metric tests as well. Equal variance is not required for the parametric one-tailed t-test used in this report, but equal variance is used for other forms of the t-test. In order to com- pare the statistical verification of equal variance with the effluent probability plots (see Section 8.5.3), Table 8-8 con- tains the results of three different variance assessments: the Cochran’s C test, the Bartlett test, and the Levene test, all included in the StatGraphics computer program. P-values that signify the correlation between influent and effluent data for each pollutant were compared with the alpha value of 0.05, and, if any of the three tests conducted calculated p- values lower than 0.05, there is a statistically significant dif- ference among the standard deviations at the 95% confidence level (StatGraphics). Not surprisingly, on the basis of the EPM plots, the hypothesis of equality of influent and effluent log EMC variance is not rejected, for all three parameters. 8.5.6 Statistical Verification of EMC Differences 8.5.6.1 Overview In order to interpret qualitative EPM plots and box plots, parametric and/or nonparametric statistical tests may be used to determine whether a significant difference exists between the influent and effluent data sets (or among EMC data for different BMPs, etc.). Parametric tests typically require that the data be normally distributed and thus have either a nor- mal or lognormal probability density function (PDF), as determined previously, and some tests may require that the variance of sample sets be the same and constant over a range of values (Burton and Pitt 2002). If parametric test require- ments are met, parametric tests should be used because they have greater statistical power (Burton and Pitt 2002). Gener- ally, nonparametric methods of determining distributions and processing data are used when normal or lognormal PDFs are determined not to be a valid method for describing the observed data because the data do not fit a particular dis- tribution function or because the sample size was too small. Nonparametric methods may be used in cases in which the frequency distribution parameters of the variable of interest are not known (Helsel and Hirsch 1992) or when the influent and effluent data sets have different distributions. Helsel and Hirsch (1992) recommend performing both parametric and nonparametric tests for small sample sizes to protect against Normality Test Comparison U.S. 183 Filter Strip Null hypothesis (H0) = data are lognormally distributed Influent Effluent Total Suspended Solids Sample Size (N) 19 23 Actual Correlation Coefficient R 0.977 0.947 Critical Correlation Coefficient R* 0.949 0.956 Alpha (α) 0.05 0.05 Shapiro Wilk P-value 0.4284 0.0085 Reject Null Hypothesis (Y/N) PPCC Normality Test N Y Shapiro Wilk Test N Y Nitrate Sample Size (N) 19 22 Actual Correlation Coefficient R 0.972 0.976 Critical Correlation Coefficient R* 0.949 0.954 0.05 0.05 Shapiro Wilk P-value 0.1698 0.225 Reject Null Hypothesis (Y/N) PPCC Normality Test N N Shapiro Wilk Test N N Total Zinc Sample Size (N) 19 23 Actual Correlation Coefficient R 0.916 0.927 Critical Correlation Coefficient R* 0.949 0.956 0.05 0.05 Shapiro Wilk P-value 0.0058 0.0012 Reject Null Hypothesis (Y/N) PPCC Normality Test Y Y Shapiro Wilk Test Y Y Alpha (α) Alpha (α) Table 8-7. Normality test comparison using the PPCC and Shapiro-Wilk tests. Variance Assessments Null hypothesis (H0) = variance of logs is the same U.S. 183 Filter Strip Total Suspended Solids 0.05 Cochran's C test (p-value) 0.606 Bartlett test (p-value) 0.607 Levene test (p-value) 0.930 Reject Null Hypothesis (Y/N) N Nitrate 0.05 Cochran's C test (p-value) 0.868 Bartlett test (p-value) 0.868 Levene test (p-value) 0.904 Reject Null Hypothesis (Y/N) N Total Zinc 0.05 Cochran's C test (p-value) 0.885 Bartlett test (p-value) 0.886 Levene test (p-value) 0.287 Reject Null Hypothesis (Y/N) N Alpha (α) Alpha (α) Alpha (α) Table 8-8. Results of the com- parison of variance testing.

71 the potential loss of power of parametric tests when distrib- utional requirements are not met. The parametric test employed for the statistical analysis of the log-transformed data is the one-tailed t-test for unequal variance. This t-test does not necessarily require equal vari- ance, although the hypothesis of equal variance of EMC loga- rithms is not rejected for the three data sets considered herein. The one-tailed t-test (as opposed to the two-tailed t-test) is deemed acceptable because the effluent concentration is pre- dicted to be less than the influent concentration for all plots. The one-tailed t-test is used to assess the difference between the means for the one constituent, total zinc, for which the influent and effluent log-transformed data sets are both nor- mal, based on the results of both the PPCC and Shapiro-Wilk tests. The nonparametric test employed to test the difference between the medians was the Kruskal-Wallis test, and this test was used on all site/constituent combinations. 8.5.6.2 One-Tailed t-test Results Results of the Excel version of the one-tailed t-test for unequal variance for the log-transformed nitrate data for U.S 183 in Austin are given in Table 8-9. The t-test assumes a null hypothesis that the means of the two groups of data are equal. It may be seen that the critical value for rejection of equality of means (p-value) is well below the predetermined value of alpha = 0.05. Hence, “removal” of nitrate by the fil- ter strip is verified by the t-test, even though removal was questionable on the basis of the scatter plot, EPM plot, and box plots. This result is also confirmed by the nonparametric test described below. 8.5.6.3 Kruskal-Wallis Test Results Nonparametric methods of data analysis typically allow for processing of more random data and assume nothing is known about the frequency distribution of the data set (Helsel and Hirsch 1992). These methods do not rely on the estimation of parameters, such as mean and standard devia- tion, to assist in describing the distribution of variables. The Kruskal-Wallis nonparametric test will be illustrated for the three data sets. The rank-sum nonparametric test (Helsel and Hirsh 1992) could also be used when just two data sets are being compared, but the Kruskal-Wallis test (suitable for comparison of two or more groups of data) is very similar and more easily available in the StatGraphics software used in this study. The null hypothesis for the analysis is that the medians of the two data sets (influent and effluent EMCs) are equal. The results of the Kruskal-Wallis test are reported in Table 8-10 for a 95% significance level (alpha = 5%). Equality of medians is rejected for all three constituents. Although equality of nitrate medians may not be rejected for many sites and BMPs, for these results, equality of medians is rejected and nitrate removal is confirmed by the Kruskal-Wal- lis test; however, the p-value for nitrate is much higher than the p-value for the other two constituents. A p-value > 0.05 would lead to acceptance of the hypothesis of equal medians and therefore no removal. The results of the Kruskal-Wallis test are consistent with the earlier t-test result as well. The bottom line for the three constituents monitored at U.S.183 in Austin and evaluated herein is that the filter strip BMP removes all three. Other nonparametric tests, such as the Hodges-Lehmann test, are similar in character to the Kruskal-Wallis test (Helsel and Hirsch 1992) and have not been pursued further herein. Other graphical techniques are also available, such as quan- tile-quantile plots (Helsel and Hirsh 1992) for comparison of frequency distributions of two different data sets. Again, the reader may access the cited literature for further information on other comparison options. 8.5.7 Simulation of EMC Removal Section 8.5 thus far has illustrated statistical methods of evaluating EMC data. However, in Section 8.7 of this chapter and Chapter10, the use of computer models to simulate high- way runoff quality and BMP performance is discussed, in particular, the USEPA SWMM model. At this point, it is appropriate to discuss briefly the most common method for Nitrate Influent Effluent Mean 0.174815 –0.52241 Variance 0.711655 0.768261 H0 = μ(x) = μ(y) Observations 19 22 Reject Null Hypothesis t Stat 2.5842 #1 t critical < t stat P(T≤t) one-tail 0.006810 #2 P(one-tail) < 0.05 t-critical one-tail 1.684875315 Table 8-9. t-test results for natural log-transformed nitrate at U.S. 183 in Austin, TX. Kruskal-Wallis Test Null hypothesis (H0) = medians of data sets are equal U.S. 183 Filter Strip Total Suspended Solids 0.05 P-value 5.60E-05 Reject Null Hypothesis (Y/N) Y Nitrate 0.05 P-value 0.010693 Reject Null Hypothesis (Y/N) Y Total Zinc 0.05 P-value 7.1E-06 Reject Null Hypothesis (Y/N) Y Alpha (α) Alpha (α) Alpha (α) Table 8-10. Kruskal-Wallis test results for difference between the medians.

72 simulation of BMP quality treatment in SWMM and how this method might be adapted to produce frequency distributions similar to those observed earlier for the U.S. 183 filter strip. This discussion is based on three reports and papers by Huber et al. (2004, 2005, 2006). Although it is not the only option available, the most straightforward method for treatment simulation in versions 4 and 5 of SWMM (Huber and Dickinson 1988; Rossman 2004) is the use of a conceptual continuous-flow, stirred-tank reactor, or CFSTR. Conceptually, the CFSTR represents all treatment or removal processes that act to reduce EMCs of constituents as they pass through the BMP (see Figure 8-14). This method does not account for beneficial losses of water by infiltration and ET, which act to reduce loads (product of flow  concentration), especially for LID facilities. However, SWMM simulation also accounts for storage changes during inflows to and outflows from the facility. The main removal mechanism available with version 5 of SWMM (http://www.epa.gov/ednnrmrl/models/swmm/index. htm) is first-order decay, for which a CFSTR conceptualization (for constant volume, to simplify the explanation) is where C(t) = effluent concentration, Co = initial concentration, k = first-order decay coefficient (1/time), and t = hydraulic detention time. In Equation 8-6, t (hydraulic detention time) is applied on a time-step basis in SWMM to reflect changes in concentra- tion as a function of dynamic changes in inflow concentra- tions and other parameters. The point is that the only option for outflow concentration is to decrease toward zero; this would mean a reduction at every time step toward zero. How- ever, as has been shown for U.S. 183 in Austin, some BMPs and constituents interact so that there is a limiting effluent EMC, or rather, a frequency distribution of effluent EMCs for which the concentrations do not approach zero and which remains relatively constant regardless of the influent EMCs. The U.S. 183 TSS and total zinc are excellent examples of this phenomenon. However, according to Equation 8-6, a CFSTR will always reduce concentrations. A solution to this dilemma recommended by Wong et al. (2002), for use in their MUSIC model and easily available in version 5 of SWMM, is to C(t) C e -6o kt = ( )− 8 employ Kadlec and Knight’s (1996) k-C∗ model for a CFSTR (The k-C∗ model is also available in the spreadsheet model of Heaney and Lee [2006], which was used extensively in this project), where Cout = effluent concentration, Cin = influent concentration, C∗ = minimum effluent EMC, k = first-order decay rate, and t = detention time. Here, the concentration tends not toward zero but toward C∗, which can represent an “irreducible minimum” effluent concentration (Minton 2005). An advantage of this formula- tion is that concentrations downstream along a treatment train or series of removal locations cannot be reduced beyond the minimum C∗. Wong (2002) reports C∗ values for TSS for “sedimentation basins” = 30 mg/L, ponds = 12 mg/L, vege- tated swales = 30 mg/L, and wetlands = 6 mg/L. Barrett (2004a, 2004b) reports values on the order of 20 mg/L for ponds and 20 to 50 mg/L for swales (the latter based largely on Caltrans studies). When the swale is designed with a filter strip component, e.g., as part of a highway embankment, pri- mary removal in the filter strip–swale combination is usually via the vegetated filter strip. However, removal in swales alone also occurs by filtration and sedimentation when the flow depth is less than the height of the vegetation. One other advantage of the k-C∗ model as an easy concep- tualization of BMP treatment is that the C∗ value may be approached both from above (dirty inflow) and below (clean inflow). When SWMM is run on a time-step basis with widely varying influent concentrations, the simulated effluent EMCs trend towards a distribution about C∗ (Huber 2006). The lim- itation of the k-C∗ model is verifying that pollutant removal follows a first-order decay and then estimating the decay coef- ficient for the particular BMP. 8.6 Overall Hydrologic and Water- Quality Performance Estimation 8.6.1 Introduction Evaluating the practicability of candidate BMPs requires consideration of several site-specific variables including expected performance for target pollutants, hydrology and hydraulics, surface and subsurface space availability, maintenance, costs, and aesthetics. Other factors include safety (see Guidelines Manual and LID Design Manual), regional constraints (see Chapter 7), and downstream C C C C e -7out in kt= − ( )+ −✱ ✱( ) 8 BMP/LID FacilityInfluent Effluent Figure 8-14. Conceptualization of BMP/LID facility treatment.

73 impacts (see Section 10.2.3). Many such variables are dis- cussed in Chapters 2 through 7. This section addresses in detail one of the most important practicability factors: treat- ment performance in terms of meeting hydrologic, hydraulic, and water-quality goals. A full practicability assessment for each candidate BMP, based on the various relevant factors, is presented in Chapter 6 of the Guidelines Manual. As many practicability factors are highly site-spe- cific, not all factors that could be relevant are included. The design engineer should build on the information presented in previous chapters to account for site-specific conditions as much as possible. Estimating the treatment performance of a BMP requires an evaluation of (1) runoff volume reductions, (2) capture efficiency, and (3) expected effluent quality for target constituents. 8.6.2 Volume Reduction There is certainly a basis for factoring in volume and result- ing pollutant load reductions into performance estimates, particularly when TMDLs are involved. The infiltration capacity of the soil within or beneath a BMP influences vol- ume reduction primarily through infiltration to subsurface and, combined with vegetation, ET. Soils with a high fraction of clays will prevent significant stormwater volume reduc- tions because of their poor infiltration capacity. If stormwa- ter volume reductions are a goal for a detention basin, soils can be amended to improve the capacity for infiltration. Higher infiltration rates will result in larger volumes entering the soils for immediate infiltration, as well as after-storm ET losses. The ET rates are also important, as they affect whether soils dry out in time to infiltrate stormwater from the next event. It is expected that wet ponds and wetland basins or chan- nels might not significantly decrease the volume of runoff because soils suitable for placement of a wet pond or wetland basin will typically exhibit low infiltration capabilities. If the soil has high infiltration capabilities, a liner will be necessary to maintain the water quality of the pool during the wet sea- son. Because of the need to maintain a permanent wet pool for optimal pollutant removal in a constructed wetland, little volume reduction can be expected because of infiltration losses. However, volume reductions would be expected in biofilters because of drier, more-permeable soils and com- plete vegetative cover. The limited study data available show an average volume reduction of 30% and 38% in dry detention basins and biofil- ters, respectively, while wet ponds (retention ponds) and wet- land basins achieve an average volume reduction of 7% and 5%, respectively (see Table 8-11) (Strecker et al. 2004a, 2004b). Based on this analysis, detention basins (dry ponds) and biofilters (vegetated swales, overland flow, etc.) appear to contribute significantly to volume reductions, even though it is likely that they are not designed specifically for this purpose. Assuming a capture efficiency (discussed in Section 8.6.3) of 80%, a dry detention basin could be expected to reduce stormwater runoff volumes by about 25% on average. The actual volume reduction depends on the infiltration characteristics of the soils and local ET rates. 8.6.3 Capture Efficiency The capture efficiency (percent of stormwater runoff treated) of an on-line, volume-based BMP (e.g., detention facility) is primarily a function of the volume of the facility and the hydraulic design of the outlet structure (e.g., brimfull or half-brimfull emptying time of the detention facility). (This concept is also discussed in Chapter 10 of this report and Chapter 7 of the Guidelines Manual.) A properly designed, storage-based BMP should generally result in capture efficiencies satisfying regulations (e.g., on the order of 70 to 90% of the long-term flows from the watershed). Untreated stormwater runoff volumes that bypass the deten- tion facility will therefore generally be less than 30% of the runoff volume, ideally on the order of 10 to 20%. For volume-based BMPs, the bypassed, untreated flows occur most often from the tail end of the storms. Depending on the pollutant and the runoff characteristics of the water- shed, these bypasses will frequently have lower pollutant con- centrations because the majority of accumulated pollutants are discharged earlier in the storm. This higher pollutant loading at the beginning of a storm event is referred to as the first-flush effect (see Sections 4.4 and 6.4). However, for many pollutants and under the varied rainfall conditions found seasonally and geographically around the United States, a first-flush effect may not exist. First flush is frequently over- simplified for the complex phenomena of pollutant source loading to runoff. Design volume (depth). In the simplified sizing approaches frequently used in regulatory environments, the design depth (inches) is the depth of runoff over the catchment that results BMP Type Mean Monitored Outflow/Mean Monitored Inflow for Events Where Inflow is Greater Than or Equal to 0.2 Watershed Inches Detention Basins 0.70 Biofilters 0.62 Media Filters 1.00 Hydrodynamic Devices 1.00 Wetland Basins 0.95 Retention Ponds 0.93 Wetland Channels 1.00 Source: Strecker et al. 2004a, 2004b. Table 8-11. Volume losses in various BMPs.

74 in a runoff volume equivalent to the storage design volume of a detention basin. The runoff volume for the tributary area is a function of the watershed size and runoff coefficient; the runoff coefficient is a function of the impervious fraction and soil type(s) in the tributary area to the basin. Larger design depths result in a larger percentage of the stormwater runoff captured by the basin. However, as design depths become excessively large, only marginal improvements are gained. Given that drain time criteria remain constant, very large design depths are undesirable because the smaller, more fre- quent storms will pass quickly through the basin without receiving sufficient detention time for sedimentation to occur (see Section 10.2 and Guidelines Manual Chapter 7). A proper design depth provides nearly complete treatment of smaller, more frequent storms and captures significant portions of larger storms. Continuous simulation screening results are provided in the Guidelines Manual, Appendix C; these results show percent capture versus design depths for 30 locations in the United States. Drawdown rate. The drawdown rate is the average outflow rate (cfs) at which a detention facility is emptied. Often the upper third of the detention basin is emptied in half of the detention time (from full pool), while the lower two-thirds is emptied in the remaining detention time. This scenario cre- ates storage capacity quickly for the next storm event, if the next storm occurs before the basin is completely empty (storm interevent times are typically within 24 hr or between 24 and 48 hr). Slower drawdown of the lower half of the pool pro- motes effective treatment for smaller storm events that would not completely fill the detention basin. Typical detention times to achieve sedimentation and removal of associated pollutants range from 24 to 72 hr. Shorter detention times (e.g., 24 hr) create storage volume more quickly, resulting in a higher cap- ture efficiency, but do not allow as much time for sedimenta- tion. An appropriate detention time should be determined on the basis of the expected particle size distribution in stormwa- ter runoff and typical storm interevent times. Longer deten- tion times are appropriate for treating runoff with a large fraction of fine particles. Shorter detention times are more appropriate for stormwater runoff with fewer fines, or in areas with storms that occur in series with short interevent times. These trade-offs are discussed in Chapter 10 of this report and Chapter 7 of the Guidelines Manual. A maximum drain time of 36 hr from a brimfull condition is often an appropriate compromise between the removal efficiency of particles and capture efficiency of stormwater runoff volumes. Drawdown at maximum is the seasonal mean time between precipitation events in a watershed. Flow rate. Flow-based treatment systems, such as most swales, are frequently sized on the basis of the calculation of a peak-flow estimate derived from a design event, unit hydro- graph, or rainfall/runoff model. One major disadvantage of flow-based design is that it does not normally account for the volume of the runoff hydrograph. A flow-based system is best sized to capture a required runoff volume (say 80%) by meth- ods presented in Section 10.3, summarized as follows: 1. Plot the historical or simulated hydrograph for its full- time base of months or years. 2. Choose a range of flow rates (represented by horizontal lines on the hydrograph). 3. Integrate along the hydrograph to find the volume beneath each flow rate (beneath each horizontal line). Convert the volume to percent of total volume under the hydrograph. This percentage is the percent of runoff volume treated by (runoff flowing through) the device at a given maximum inflow rate. 4. Plot flow rate versus percent runoff volume captured, as in Figure 10-17. The plot yields the particular inflow rate necessary to treat a specified percent runoff volume (e.g., 80%) and vice versa. Based on this runoff hydrograph analysis, the flow-based sys- tem could then be sized for the flow rate that would capture the runoff volume to be treated. Flow-based treatment sys- tems are evaluated in Appendix C of the Guidelines Manual for 30 locations in the United States. 8.6.4 Pollutant Removal 8.6.4.1 Summary Data Median effluent quality for various BMPs is shown in Table 8-12 for common target constituents. The data are from the International BMP Database (www.bmpdata base.org). Data summaries by typical stormwater pollutants are also provided in the pollutant fact sheets (see Appendix A of the Guidelines Manual). The degree of pollutant removal, of course, depends on the pollutant species/form and the extent to which appropriate unit operations occur within the treatment system. In addition, design features such as pond depth or use of a forebay can significantly affect effluent quality. 8.6.4.2 Suspended Solids Larger suspended solids can be removed effectively by gravitational sedimentation, screening, or surficial straining. For most well-designed BMPs that incorporate these unit operations, the median effluent concentrations range from 20 to 25 mg/L, provided the concentration and characteristics (e.g., particle size distributions) of influent suspended solids

75 do not significantly deviate from typical stormwater. Well- designed treatment systems that incorporate wet pools and wetland vegetation typically exhibit good effluent quality for suspended solids. Available data suggests that these BMPs can typically achieve effluent concentrations of around 20 mg/L. Well-designed biofilters and media filters also perform well in achieving low concentrations of effluent suspended solids. The presence of a permanent wet pool is a feature of a wet pond/wetland system. Incorporating even a small permanent wet pool can significantly improve the sediment removal per- formance of a BMP by providing long periods of retention during smaller storms. Long retention times during small events allow for appreciably more sediment removal than dry facilities that typically have very limited detention times dur- ing small events. Generally, settleable solids composed of inorganic particles in the 25- to 75-μm range are effectively removed by quiescent gravitational sedimentation. For biofilters and media filters, gravity settling and fil- tration are the primary removal mechanisms for suspended sediments. Direct filtration can usually be effectively accomplished at concentrations less than 50 mg/L, but gen- erally requires some level of pretreatment in urban runoff, in which solids concentrations are frequently above 100 mg/L and can exceed 1,000 mg/L depending on the site, loading, and hydrology. Generally, suspended inorganic particles less than 25 μm require some natural or enhanced coagulation/flocculation followed by sedimentation and/or filtration. Available data suggest that TSS effluent concentrations are significantly higher (i.e., the quality of the effluent is poorer) for dry detention basins (which drain after each event and generally lack a significant littoral zone) and hydrodynamic BMPs (flow-through systems that rely on centrifugal forces to provide treatment). However, as noted, dry detention basins Constituents Detention Pond Wet Pond Wetland Basin Biofilter Media Filter Hydrodynamic Devices Suspended Solids 41.35* (30.8–55.5) 19 (12.9– 28.0) 19.68 (16.6–23.4) 24.6 (15.0–40.3) 25.47 (14.7– 44.3) 40.34 (18.4–88.7) Total Cadmium 1.3 (0.8–2.2) 0.31 (0.05– 2.0) xx** 0.25 (0.21–0.34) 0.31 (0.16– 0.59) 1.65 (1.05–2.6) Dissolved Cadmium 0.41 (0.22–0.76) xx xx 0.22 (0.11–0.43) 0.24 (0.18– 0.33) 0.93 (0.27–3.2) Total Copper 18.9 (16.6–21.5) 6.92 (4.7– 10.3) xx 10.01 (5.6–17.9) 9.81 (8.1– 11.8) 14.13 (11.1–18.1) Dissolved Copper 14.72 (10.4–20.9) 5.09 (3.1–8.3) xx 7.66 (4.7–12.5) 7.95 (6.6–9.7) 8.63 (3.3–22.9) Total Chromium 2.85 (1.7–4.8) 1.78 (0.5-6.7) xx 2.18 (1.2–4.0) 1.46 (0.9–2.3) xx Total Lead 15.02 (9.5–23.8) 6.68 (2.9– 15.6) 3.25 (1.9–5.6) 6.95 (4.2–11.7) 5.5 (3.5–8.6) 12.98 (4.2–40.2) Dissolved Lead 2.33 (1.7–3.3) 4.16 (2.0–8.9) xx 1.35 (0.5–3.6) 1.42 (1.0–1.9) 2 (0.6–6.5) Total Zinc 85.26 (50.6–143.7) 28.63 (21.4– 38.3) 118.73 (32.8–429.5) 39.44 (28.2–55.2) 64.96 (45.3– 93.2) 89.66 (74.4–108.1) Dissolved Zinc 43.99 (20.0–96.6) 16.89 (2.6– 109) xx 31.96 (26.7–38.3) 57.14 (37.7– 86.6) 45.17 (29.6–68.9) Total Phosphorus 0.3 (0.2–0.44) 0.16 (0.12– 0.21) 0.15 (0.07–0.33) 0.32 (0.24–0.43) 0.14 (0.11– 0.17) 0.19 (0.07–0.51) Dissolved Phosphorus xx 0.07 (0.04– 0.13) 0.07 (0.03–0.18) xx xx xx Total Nitrogen xx 1.17 (0.77– 1.78) 2.42 (1.46–4.0) 0.69 (0.37–1.29) xx xx Nitrate- Nitrogen 0.64 (0.37–1.09) 0.48 (0.11– 2.05) 0.46 (0.16–1.28) 0.5 (0.36–0.68) 0.82 (0.68– 0.97) xx Total Kjeldahl Nitrogen 1.87 (1.46–2.39) 0.84 (0.68– 1.04) 1.33 (0.84–2.11) 1.6 (1.42–1.8) 1.79 (1.45– 2.2) 4.68 (1.97–11.12) * All units in mg/L; values in parentheses are the 95% confidence intervals about the median. ** xx = lack of sufficient data to report median and range. Source: International Stormwater BMP Database (http://www.bmpdatabase.org/). Table 8-12. Median of average effluent concentrations of BMPs.

76 have been shown to provide considerable reduction in efflu- ent volume (up to 30%), which may translate into lower total mass loading of TSS downstream. 8.6.4.3 Trace Metals The important forms of trace metals from a treatability and regulatory perspective are total, dissolved, and particu- late-bound metals. If trace metals are bound to organic or inorganic particulates, viable unit operations include sedi- mentation and filtration either as unit operations separate from coagulation/flocculation or in combination with coag- ulation/flocculation as pretreatment to these operations. If trace metals are present as a dissolved complex, precipitation could be effective. If trace metals are present as a dissolved ionic species such as Cu2+, Pb2+, or Zn2+, surface complexation (including adsorption) could be effective. Well-designed wet ponds, biofilters, and media filters can provide better effluent quality than detention ponds and hydrodynamic devices (see Table 8-12). BMPs that are effective in removing trace metals also are typically good at removing fine particulates. 8.6.4.4 Nutrients Treatability for phosphorus is a function of whether phos- phorus is present in particulate or dissolved form. In dis- solved form, phosphorus may readily undergo surface complexation reactions, sorption, or precipitation (see Sec- tion 4.4.5). Uptake by vegetation and microbes is another mode by which dissolved phosphorus is effectively removed. Media or soils containing iron, aluminum, or hydrated portland cement can be very effective at removing dissolved phosphorus species through surface complexation or precip- itation. If phosphorus is bound to organic or inorganic par- ticles, viable unit operations include sedimentation and filtration either alone or in combination with pretreatment using coagulation/flocculation. As shown in Table 8-12, media filters, wet ponds, and wet- land basins report the lowest median effluent concentrations of total phosphorus, although only wet ponds show a statis- tically significant difference between median influent and effluent values (i.e., the BMP affected total phosphorus con- centrations). Although median effluent levels for dissolved phosphorus are the lowest for wetland basins, the available data are insufficient to reliably differentiate the performance of various BMPs. Nitrogen compounds exist in dissolved form and as particulate-bound species. Treatability success for nitrogen species, as with other constituents in stormwater, is highly dependent on the form and species of nitrogen present. Treatability for nitrogen also depends on the presence of specific bacteria that mediate nitrogen transformations. Physical operations such as sedimentation have played an insignificant role with respect to treatment of nitrogen as compared with microbially mediated transformations. Microbial decomposition of organic matter mineralizes nitrogen as ammonia, which can be oxidized to nitrite and nitrate. Nitrate can be reduced to nitrogen gas by anaerobic bacteria for complete removal from the system. Median effluent quality of total nitrogen, TKN, and nitrate-nitrogen are summarized in Table 8-12. However, available data on removal of nitrogen species are insufficient to draw defini- tive conclusions about BMP performance based on average effluent concentrations. Filters, ditches, and dry ponds typically exhibit poor nitrate removal and, in many cases, have been shown to export nitrate. In these BMPs, organic nitrogen is converted to nitrate in the mineralization and nitrification processes; how- ever, the aerobic conditions are not favorable for denitrifica- tion. Thus, these BMPs may export more nitrate than is present in the influent. Conversely, in wet ponds and wetland basins, plants, algae, and other microorganisms take up nitrate as an essential nutrient. However, nitrogen is also released back into the system upon death or decay of the organisms. 8.7 Methodology Options Using Process Simulation Models Simulation models provide an opportunity to analyze details of several BMP options, including treatment trains and the ability of some BMPs to evapotranspire and/or infil- trate runoff. For example, the Storage/Treatment Block of version 4 of SWMM (Huber and Dickinson 1988; Roesner et al. 1988) contains a process-oriented approach for evalu- ating the effectiveness of BMPs. (This is demonstrated in Chapter 7 of the Guidelines Manual and discussed in Section 8.5.7.) All treatment systems employ some mix of physical, chemical, and biological processes to achieve some effluent quality and/or removal efficiencies. For wet-weather con- trols, many of the BMPs include temporary storage of stormwater. When stormwater is stored temporarily, it is vital to include in the analysis the dynamics of the filling and emptying of these devices because these have a major impact on the amount of runoff treated and removal efficiencies. For example, detention time is difficult to estimate because it depends on the mixing regime (e.g., CFSTR or plug flow) and how the outlet control is operated (e.g., drawdown regime). In addition, some of the stormwater detained in a pond in a warm, arid area of the country may evaporate before the next storm event, thereby increasing the overall performance of the BMP. An infiltration system may receive irrigation water between storm events, thereby reducing its ability to infiltrate runoff during the next storm event.

77 Simulation models also allow one to evaluate a long-term rainfall record to assess how much stormwater is treated by the BMP and how much is bypassed or processed by the BMP at the lower, water-quality design flow rate, and the higher, flood control rate at which higher flows (and result- ing short detention times) will result in minimal or ineffec- tive treatment. For example, swales are not effective at flow rates above a certain level (often the flow rate that results in a flow depth which covers the vegetation). Any flows with rates above the design flow rate should be considered as a “bypass” during a BMP performance analysis. Thus, evalua- tions of the performance of a BMP must take into consider- ation how much of the rainfall record is treated, controlled, or eliminated. An additional option for simulation of the rainfall record for these kinds of evaluations is the spread- sheet model of Heaney and Lee (2006). Various SWMM components allow the user to perform continuous simulations over extended periods of time in order to integrate treatment process and hydraulic dynamics. Heaney and Nix (1979) summarize the basic ideas behind this approach. Medina et al. (1981a, 1981b) present a more complete description of the process dynamics for doing con- tinuous simulation. Nix (1982) presents a comprehensive evaluation of statistical and process-oriented approaches for evaluating storage-release systems. Goforth et al.(1983) com- pare a process simulator with a statistical approach for esti- mating BMP performance. Nix and Heaney (1988) describe how to optimize the size and release rate of stormwater con- trol devices. Finally, Nix et al. (1988) describe a more detailed process approach for evaluating suspended solids removal by relating it to particle sizes and detention times. In addition to using the Storage/Treatment Block of version 4 of SWMM, Huber (2001) and Huber et al. (2006) show how the per- formance of many of the BMPs can be simulated using the Runoff Block of version 4 of SWMM. The newest version of SWMM is version 5, which was developed with a graphical user interface by the USEPA (www.epa.gov/ednnrmrl/mod els/swmm/index.htm; Rossman 2004). In general, version 5 of SWMM is much easier to use than version 4; however, ver- sion 5 lacks some version 4 functionality, specifically the abil- ity to simulate sedimentation using sedimentation theory with a plug flow assumption. Versions 4 and 5 of SWMM applications are outlined in Chapter 10 of this report and pre- sented in detail in Chapter 7 and Appendix E of the Guidelines Manual. Continuous simulation modeling is an important tool in assessing BMP performance. A number of efforts have improperly applied flood design hydrology approaches to sub- stantiate hydraulic performance (Strecker et al. 2001). SWMM is far from the only continuous simulation tool available; the Hydrologic Simulation Program—Fortran (HSPF) (Bicknell et al. 1997) is an example of a very widely used model for sim- ulation of runoff and water quality from rural and urban watersheds. Another widely used model (for hydrology only) is the well-known HEC-HMS model (www.hec.usace.army. mil/), which can also be used for long-term, period-of-record (i.e., continuous) simulation. Model choice (also discussed in Section 10.5.2) is based on model capabilities and often on user familiarity. Both factors led to the use of SWMM in this project.