Download as pdf or txt
Download as pdf or txt
You are on page 1of 24

Water Air Soil Pollut (2012) 223:2719–2742

DOI 10.1007/s11270-011-1062-8

A System Dynamic Model and Sensitivity Analysis


for Simulating Domestic Pollution Removal in a Free-Water
Surface Constructed Wetland
Yung-Chieh Wang & Yu-Pin Lin &
Chun-Wei Huang & Li-Chi Chiang &
Hone-Jay Chu & Wen-Sheng Ou

Received: 11 July 2011 / Accepted: 20 December 2011 / Published online: 7 January 2012
# Springer Science+Business Media B.V. 2012

Abstract This work develops a system dynamic simu- (DO), 5-day biological oxygen demand (BOD5), total
lation model for free-water surface constructed wet- nitrogen (TN), total suspended sediment (TSS), and
lands, as well as provides appropriate values for the total phosphorous (TP) concentrations, the model can
parameters of constructed wetland management. The simulate the variations of DO, BOD5, and TSS. Taken
system dynamic model is calibrated and validated by into account the interactions among parameters, the
using data from a 1-year study of a constructed wetland GLUE method successfully obtained the model sensi-
in Tainan of southern Taiwan. Additionally, the major tive parameters from the Monte Carlo parameter sets.
parameters that affect the simulation output are obtained Sensitivity analysis results indicate that the parameters
via sensitivity analysis by using generalized likelihood of microorganisms are sensitive factors that affect DO,
uncertainty estimation (GLUE). A high R2 and Nash– BOD5, and TN, while sediment diameter largely influ-
Sutcliffe coefficient of efficiency between the simulated ences TP and TSS. Further elucidating environmental
and measured outflow values indicate that in addition to microorganisms would increase the model accuracy and
reproducing the changing trends of dissolved oxygen provide a valuable reference for constructed wetland
management and design.

Y.-C. Wang Keywords Constructed wetland . Water quality .


School of Civil and Environmental Engineering,
System dynamic model . Sensitivity analysis
Georgia Institute of Technology,
Atlanta, GA 30332, USA

Y.-P. Lin (*) : C.-W. Huang : L.-C. Chiang 1 Introduction


Department of Bioenvironmental Systems Engineering,
National Taiwan University,
Taipei 10617 Taiwan, Republic of China Free-water surface (FWS) constructed wetlands (CWs),
e-mail: yplin@ntu.edu.tw have been widely applied at various stages of wastewater
treatment and to treat water from different sources and
H.-J. Chu
with diverse characteristics (Ghermandi et al. 2007;
Department of Geomatics,
National Cheng Kung University, Vymazal 2007). In FWS CWs, the effluent quantity of
Tainan 70101 Taiwan, Republic of China suspended particles is usually reduced physically by ag-
gregation, sedimentation, filtration, and adhesion onto
W.-S. Ou
macrophytes; chemical processes include volatilization
Department of Landscape Architecture,
National Chin-Yi University of Technology, of ammonium, as well as adsorption and ion exchange
Taichung 41170 Taiwan, Republic of China of phosphorous and heavy metals (Merz 2000).
2720 Water Air Soil Pollut (2012) 223:2719–2742

Microorganisms (both autotrophic and heterotrophic) and Dennett 2008; Wang et al. 2009); thus, a dynamic sim-
aquatic macrophytes both play important roles on the ulation model which describes most of the major nutrient
decomposition, mineralization, transformation (nitrifica- cycles in FWS constructed wetlands is expected.
tion and denitrification), and biological uptake from water Sensitivity and uncertainty analyses are primarily
column (Kadlec and Knight 1996; Merz 2000; Wynn and concerned with the question of how model outputs are
Liehr 2001). Natural filtration through FWS systems affected by the variability of the model parameters and
provides a water reclamation solution that can meet reuse input values. Sensitivity analysis is the determination
quality standards for a variety of applications, in particu- of which parameters predominately control the model
lar for no contact to medium quality uses (Ghermandi et behavior (Beven and Freer 2001; Rankinen et al.
al. 2007). The removal performance of conventional 2006; Zhang et al. 2009), whereas uncertainty analysis
wastewater treatment parameters such as the biochemical is the estimation of error in the model output due to
and chemical oxygen demand (biological oxygen de- uncertainty in the model structure, parameters and data
mand (BOD), COD), total suspended solids (TSS), and inputs (Rankinen et al. 2006; Dean et al. 2009). Gen-
nutrients has been object of extensive research in the last eralized likelihood uncertainty estimation (GLUE) is
two decades (Kadlec et al. 2000; Ghermandi et al. 2007). an extension of the regional sensitivity analysis tech-
As the popularity of, and interest, in wetlands have nique proposed by Spear and Hornberger in 1980 (Dean
increased, so has interest in modeling the processes that et al. 2009). The GLUE technique developed by Beven
occur within a wetland. However, a comprehensive and Binley (1992) has been recently used extensively
understanding of these processes is desirable in order for simultaneous calibration and uncertainty assessment
to design, model, manage, monitor, and maintain con- of hydrological and water quality models (Pastres and
structed wetlands. Different types of models such as Ciavatta 2005; Mantovan and Todini 2006; Rankinen et
simple first-order plug flow model, K-C model, regres- al. 2006; Pohlert et al. 2007). Unlike other methods, the
sion model, hydrodynamic model, stochastic model, GLUE technique allows assessment of the global uncer-
and system dynamic model have been developed to tainty present in various modeling elements (e.g., input
simulate the processes of nutrient cycles in wetlands data, model structure, and parameter errors) in a way
(Kadlec and Knight 1996; Lee 1999; Wynn and Liehr that is conceptually easy to implement (Blasone et al.
2001; Mayo and Bigambo 2005; Marsili-Libelli and 2008).
Checchi 2005; Chavan and Dennett 2008). Each type The objective of this study is to develop a simulation
of model has its own potential and constraints with model for FWS constructed wetlands by using SD soft-
respect to the requirements and expectations of simulat- ware STELLA 5 (High Performance Systems 1990–
ing nutrient cycles. Moreover, these wetland models 1997) in order to realize the interactions between circu-
range from simple simulation using the wetland hydro- lations of nutrients in wetland systems. From January to
logical cycle to more complex models involving hydro- December 2004, weekly sampling data were collected
logical, N, P, TSS, and vegetation cycles (Kadlec and from an FWS-constructed wetland pond in National
Knight 1996; Lee 1999; Wynn and Liehr 2001; Lee et Cheng-Kung University, Tainan, Taiwan. Furthermore,
al. 2002; Mayo and Bigambo 2005; Marsili-Libelli sensitivity analysis using the GLUE approach was ap-
and Checchi 2005; Chavan and Dennett 2008). plied to find the important parameters that affect water
System dynamics (SD) approach, created in the mid- treatment efficiency, and then calibration and validation
1950s, is a powerful methodology and computer simu- were used to assess the accuracy of the model. Accu-
lation modeling technique for framing, understanding, rate prediction of system dynamics and identifica-
and discussing complex issues and problems and has tion of the major parameters that affect FWS
been used to model wetland systems by a number of constructed wetland treatment efficiency would pro-
researchers, e.g., Lee (1999), Wynn and Liehr (2001), vide useful tools for the design and operation.
Mayo and Bigambo (2005), Hafner and Jewell (2006),
Chavan and Dennett (2008), and Wang et al. (2009).
However, many of the simulation models for constructed 2 Material and Methods
wetlands focus on either a few nutrient cycles or subsur-
face flow systems (Wynn and Liehr 2001; Mayo and The model was developed using software STELLA 5
Bigambo 2005; Hafner and Jewell 2006; Chavan and following the structural procedure for model development.
Water Air Soil Pollut (2012) 223:2719–2742 2721

The physical, chemical, and biological processes in the would be a multi-objective problem. Assuming equal
proposed FWS constructed wetland model followed weights for each objective, we integrate the multiple
the descriptions in several studies (Lee 1999; Wynn objectives (Nash–Sutcliffe coefficient of efficiency
and Liehr 2001; Mayo and Bigambo 2005; Chavan (NSE) values of all target outputs, including DO,
and Dennett 2008). Figure 1 show a brief overview of BOD5, TN, TSS, and TP) into a single objective
the parameter subset selection and tuning procedure function. Therefore, maximum of total Nash–Sutcliffe
referred the procedure of Brun et al. (2002) and coefficient of efficiency, NSEtotal would be represented
Chavan and Dennett (2008). In this study, model as the objective function in this study (step 2). A total
calibration was included in the procedure. Next, of 32 parameters were evaluated in this study (Table 1).
single-sensitivity analysis and uncertainty analysis of The initial values of all parameters were given by
model parameters were then processed. For model referring to previous studies or the mean of the range
calibration (from step 1 to 6), we have to define the of literature values (step 3). One thousand parameter
physical, chemical, and biological mechanism in the set within the range of +10% and −10% of the initial
proposed FWS constructed wetland model in the first values of the current parameter set is then generated
step. Because of the influences between parameters by Monte Carlo simulation (step 4). After ranking the
and five outflow concentration predictions (dissolved performance of all parameter set, the current parameter
oxygen (DO), 5-day biological oxygen demand set would be replaced by better one (step 5). The
(BOD5), total nitrogen (TN), total suspended sediment iteration to obtain the best set of parameters is stopped
(TSS), and total phosphorous (TP)), model calibration until each NSE value of target outputs is satisfied with

Fig. 1 Procedure of a
systematic parameter set
selection
2722 Water Air Soil Pollut (2012) 223:2719–2742

Table 1 Model calibration parameter values

Parameter Description Literature value Source Calibration

Autotroph system
δNSa Death rate of Nitrosomonas (1/day) 0.002 Wynn and Liehr (2001) 0.002
μmax_NSa The maximum growth rate of Nitrosomonas 0.01 Wynn and Liehr (2001) 0.0108
(1/day) 0.33–2.21 Jorgensen et al. (1991); Mayo
and Bigambo (2005); Wang et
al. (2009)
KNH4_NSa NH4+-N half saturation constant of NS (g/m3) 0.8 Weber and Tchobanoglous (1986) 1.24
1.0 Wynn and Liehr (2001)
KDO_NSa DO half saturation constant of NS (g/m3) 0.15–2.0 Kadlec and Knight (1996) 0.8378
1.0 Wynn and Liehr (2001)
Heterotroph system
δ`a Death rate of heterotroph (1/day) 0.005 Wynn and Liehr (2001) 0.005
μmax_aHTa The maximum growth rate of aerobic 0.015 Wynn and Liehr (2001) 0.018
heterotroph (1/day) 0.58; 3.65 Mashauri and Kayombo (2002)
μmax_anHTa The maximum growth rate of anaerobic 0.014 Wynn and Liehr (2001) 0.018
heterotroph (1/day) 0.58; 3.65 Mashauri and Kayombo (2002)
KNO3_HTa NH4+-N half saturation constant of HT (g/m3) 0.15 Wynn and Liehr (2001) 0.1887
KDO_HTa DO half saturation constant of HT (g/m3) 1.0 Wynn and Liehr (2001) 1.0
0.1–1.0
0.08; 0.9 Mashauri and Kayombo (2002)
KTOC_HTa TOC half saturation constant of HT (g/m3) 50 Wynn and Liehr (2001) 32.035
16.2232; Marsili-Libelli and Checchi (2005)
64.492
55; 185 Mashauri and Kayombo (2002)
Dissolved oxygen system
YHT_DOa DO yield coefficient of HT (g microbes/g O2) 0.813 Wynn and Liehr (2001) 6.22
YNS_DOa DO yield coefficient of NS (g microbes/g O2) 0.084 Wynn and Liehr (2001) 11.17
Phosphorous System
PTP Fraction of P in plant biomass (g P/g biomass) 0.0008 Jorgensen (1983) 0.0008
PPmina PP mineralization coefficient (1/day) N/A N/A 0.3
FacP/SSa The surface area fraction of SS occupied by P N/A N/A 0.1309
Carbon System
PTc Fraction of C in plant biomass (g C/g biomass) 0.47 Wynn and Liehr (2001) 0.4
PTdegration_r Biomass degradation rate of plants (1/day) 0.0063 Wynn and Liehr (2001) 0.0063
PTgrowth_ra Biomass growth rate of plants (1/day) 0.059TN+ Polomski et al. (2008) 0.059 TN+
13.48 13.48
BODCa Fraction of C in influent BOD (g C/g BOD5) 0.82 Wynn and Liehr (2001) 0.725
DOCleaching_r DOC leaching rate (1/day) 0.15 Kadlec (1986) 0.15
PCaccumulate_ra Peat carbon accumulate rate (1/day) 2300 Wynn and Liehr (2001) 2300
PCa Carbon fraction of peat (g C/g peat) 0.8 Wynn and Liehr (2001) 0.8
SSCa Carbon fraction of suspended sediment (g C/g SS) N/A N/A 0.8
MicrobNa Nitrogen fraction of microbes (g N/g microbes) 0.124 Wynn and Liehr (2001) 0.124
YHTa Heterotroph yield coefficient (g microbes/g C) 0.34; 0.75 Mashauri and Kayombo (2002) 1.0
Nitrogen system
PTC:N Ration of C to N in plant biomass 23.5 Wynn and Liehr (2001) 23.5
(g C/g N)
Water Air Soil Pollut (2012) 223:2719–2742 2723

Table 1 (continued)

Parameter Description Literature value Source Calibration

Dr20a Denitrification rate constant at 20°C 0.0–1.0 Baca and Arnett (1976) 3.74
(1/day) 1.1–1.2 Tchobanoglous and Burton (1991)
0.025–0.61 Kadlec and Knight (1996)
0.002–0.46 Arheimer and Wittgren (2002)
0.195 Mayo and Bigambo (2005)
0.09–0.89 Chavan and Dennett (2008)
0.1369; 0.2197 Sonavane and Munavalli (2009)
θ1a Arrhenius constant for denitrification 1.02–1.09 Baca and Arnett (1976) 1.0485
1.08–1.10 Tchobanoglous and Burton (1991)
1.09 Beutel et al. (2009)
DONfixation_ra Fixation rate of DON (1/day) N/A N/A 0.2
YNSa Nitrosomonas yield coefficient 0.03–0.13 Charley et al. (1980) 0.02915
(g microbes/g N) 0.15 Mayo and Bigambo (2005); Wang
et al. (2009)
PN Nitrogen fraction of peat (g N/g peat) 0.025 Wynn and Liehr (2001) 0.025
SSN Nitrogen fraction of suspended sediment N/A N/A 0.2
(g N/g SS)
Suspended sediment system
dssa Settling particle diameter (m) 1×10−8 −1× Kadlec and Knight (1996) 4.91×10−6
10−6 (for clay)
1×10−6 −1× Kadlec and Knight (1996)
10−5 (for silt)
5×10−6 Mayo and Bigambo (2005)
a
Sensitive parameters for GLUE analysis (i.e. parameters which δ are greater than 0.1%)

the stop criterion (i.e., NSEDO >0.5) to ensure the 2.1 Study Area and Data
accuracy of all the model output predictions (step 6).
R2 and the NSE values of simulated output against The study area is a wetland pond constructed by the
measured outflow concentrations are applied for both Department of Architecture in National Cheng-Kung
model calibration and validation. Single sensitivity University. The FWS pond treats wastewater from the
analysis is then performed to obtain the sensitive department building. The FWS treatment system, where
parameters (step 7) and produce parameter important the data was collected, is 18.0 m long, 2.0 m wide, and
ranking (step 8) according to the predictions of the 0.8 m deep with a 0.3 m layer of soil on the impermeable
system dynamic model to identify the sensitive param- base. The front half of the pond is planted with cattail
eter subset (step 9). Finally, uniform sampling is used (Typha orientalis Presl.) and the rear half is planted with
within the specified ranges that will normally produce water hyacinths (Eichhornia crassipes). During the
a good correlation with measured data (Page et al. study period (January–December 2004), the water vol-
2003). Moreover, 5,000 parameter sets with the +30% ume was maintained at approximately 12.2 m3 with 0.85
and −30% of above sensitive parameters from the porosity (water depth 0.4 m). The measured wastewater
single-parameter analysis were applied in GLUE uncer- inflow rate was 2 m3/day from January to February 2004
tainty analysis (step 10). In addition, the measured and and 3 m3/day from March to December 2004. The
simulated outflow values are statistically analyzed using retention time is the volume of the reactor divided by
Statistical Package for Social Science 10.0 (SPSS Inc. the flow rate. Thus, the hydraulic retention time was 6.1
1989–2006). per day during the first period and 4.1 per day in the
2724 Water Air Soil Pollut (2012) 223:2719–2742

second period. Further details of the FWS system’s Jewell 2006; Wang et al. 2009). While many wetlands
operations can be found in Ou et al. (2006). The data are simulated as plug flow reactors (such as first-order
was collected weekly from January to December 2004 models: Kadlec and Knight 1996; Diederik et al. 2004;
(Ou et al. 2006). For each water sample, the concen- Chavan and Dennett 2008; Chavan et al. 2008), the
trations of DO, BOD5, ammonium-nitrogen (NH4+), actual flow conditions are between plug flow and
nitrate (NO3−), organic nitrogen (ON), TP, and TSS completely mixed scenarios; thus, a continuously
were analyzed for both the inflow and the outflow. All stirred tank reactor may be an alternative approach
the analysis samples and data were collected according (Lee 1999; Reed et al. 1995; Wynn and Liehr 2001).
to the standard method (APHA et al. 1995). The data Because of free lying water in the FWS systems, two
collected between January and June were used for mod- pools of water (a free-surface water layer and a sub-
el calibration, and those collected between July and strate water layer) are simulated for such systems (Lee
December were used for model validation. Climatolog- 1999). Accordingly, the state variables with similar
ical parameters, such as the temperature, daily precip- circulation processes in each submodel are divided
etation, and day length were obtained from Taiwan’s into two categories for the free-surface and substrate
Central Weather Bureau. Parameters that were not mea- water layers.
sured in the study area, such as the maximum growth
rates, half-saturation coefficients, and death rates of 2.2.1 Hydrological Submodel
microbes and macrophytes, were derived from previous
studies (Dawson and Murphy 1972; Jorgensen 1983; The hydrological submodel calculated the water mass
Lee 1999; Wynn and Liehr 2001; Mayo and Bigambo balance of the hydrological cycle by the dynamic
2005; Polomski et al. 2008) and recalibrated by our water budget approach (Kadlec and Knight 1996;
model. Lee 1999; Chavan and Dennett 2008; Hull et al.
2008). The input data, including the inflow, outflow,
2.2 Model Construction precipitation, and air temperature, were obtained from
the field investigation. The water lost through evapo-
The proposed FWS model is comprised of eight sub- transipiration was estimated by Thornthwaite’s meth-
models representing the water budget, dissolved oxy- od using the following equation:
gen balance, carbon cycle, nitrogen cycle, suspended
nH nD 10  T a
particle content, phosphorous cycle, and bacteria pop- PET ¼ 1:6  ð Þð Þð Þ ; ð1Þ
ulations (autotrophic and heterotrophic) in wetland 12 30 I
systems (Fig. 2). where, PET0potential evapotransipiration (centime-
In the model, it is assumed that the wetland acts as a ter), nH0the number of daylight hours, nD0the num-
single continuously stirred tank reactor (Lee 1999; ber of days in the month, T0the average monthly air
Reed et al. 1995; Wynn and Liehr 2001; Hafner and temperature (°C), and I0the annual heat index. The
heat index, I, and coefficient, a, were calculated using
the equations proposed by Jorgensen (1983).

2.2.2 Dissolved Oxygen Submodel

In the oxygen submodel, the sediment’s oxygen de-


mand, which is the oxygen consumed in the substrate
layer, is represented by oxidation processed by auto-
trophic and heterotrophic bacteria including their res-
piration patterns (Lee 1999; Wynn and Liehr 2001).
Biomass flux is the product of biomass oxygenation
resulting from photosynthesis. The oxygen is released
from the root zone of macrophytes in wetlands (Merz
2000; Hull et al. 2008; Kadlec 2008). Based on the
Fig. 2 Conceptual model of a FWS constructed wetland theory that processes like Knudsen diffusion are
Water Air Soil Pollut (2012) 223:2719–2742 2725

responsible for root zone aeration (Grosse 1991), it is slope of the wetland (meter per meter), and T0water
assumed that there is a uniform vegetation stand temperature (degree Celcius).
throughout the wetland system; thus, plants transport
oxygen to their roots at a constant rate during the 2.2.3 Carbon Submodel
growing season. Autotrophic and heterotrophic respi-
ration patterns were calculated by dividing the micro- In the carbon submodel, the particulate organic carbon
bial growth by the oxygen yield rate of microbes (Lee (POC) is defined as a homogeneous mixture of decay-
1999; Wynn and Liehr 2001). ing plant litter, sloughed microbial cells, and particu-
The diffusion of DO between surface and substrate late influent wastewater BOD (Jansson and Persson
pore water was estimated by the following mass trans- 1982). The dissolved organic carbon (DOC) is a ho-
fer equation (Lee 1999): mogeneous mixture of dissolved influent BOD and
carbon compounds leached from decaying plant ma-
DOMT ¼ KDO MT  ðCBottom DO  CSurface DO Þ  A; terial. Biomass carbon is increased by plant growth
ð2Þ and reduced by plant death and the harvesting of live
plants. The amounts of growth, death, and removal of
where, DOMT 0the diffusion of DO from one pool to
living plants are determined by the rates of plant
the other (gram per day), KDO_MT 0mass transfer co-
growth and death, as well as the removal of parts of
efficient of DO (one per day), Csurface_DO 0the DO
living plant parts; the amounts vary according to the
concentration in surface water (gram per cubic meter);
season. Standing dead carbon as the results of plant
CBottom_DO 0the DO concentration in substrate pore wa-
death is reduced by the removal of dead plant parts,
ter (gram per cubic meter), and A0the area of water–
physical degradation, and leaching of standing dead
substrate interface (square meter).
carbon to DOC.
The reaeration process in wetlands is usually
POC in surface and substrate pore water is increased
explained by the general two-film theory, which is
by POC inflow, physical degradation of plants, and
based on mass transfer (Jorgensen 1983). The follow-
microbial death; and reduced by POC mineralization,
ing equations have been used to model the reaeration
peat accumulation, and POC outflow. POC exchange
process (Lee 1999; Wynn and Liehr 2001):
between the free surface and substrate water layers
DOreair ¼ KaðT  CÞ  Adif  ðCs DO  CSurface DO Þ; occurs through settling and resuspension of POC. The
ð3Þ two processes are related to sediment deposition and
resuspension modeled by the suspended sediment sub-
where, model. DOC in surface and substrate pore water is in-
creased by DOC inflow and dead biomass leaching, and
KaðT  CÞ ¼ Kað20 CÞ θðrT20Þ ; ð4Þ
reduced by DOC mineralization and DOC outflow. The
mass transfer of DOC between free surface and substrate
pore water through diffusion was estimated according to
Kað20 CÞ ¼ 0:116 þ 2147:8S 1:2 ; ð5Þ the concentration gradient between the two layers.
The mineralization of POC and DOC, combined
with mineralization, immobilization, and ammonifica-
Cs DO ¼ 14:61996  0:40420  T þ 0:00842 tion of particulate and dissolved organic nitrogen
 T 2  0:00009  T 3 ; ð6Þ (PON and DON), are modeled by an estimating meth-
od based on the works of Gidley (1995), Lee (1999),
where, DOre-air 0the oxygen added through reaeration and Wynn and Liehr (2001). Because of the presence
(gram per day), Ka(T°C) 0the reaeration coefficient at of anaerobic microsites in high DO concentration
T°C (one per day), Adif 0the surface area through areas and aerobic rootzones in low DO concentration
which diffusion takes place (square meter), Cs_DO 0 areas, anaerobic and aerobic bacteria are always pres-
the oxygen saturation concentration (gram per cubic ent. It is assumed that the percentage of each bacteria
meter), CSurface_DO 0the DO concentration in the water and the average heterotrophic yield are functions of
layer (gram per cubic meter), Ka(20°C) 0the reaeration DO concentration (Tchobanoglous and Burton 1991;
coefficient at 20°C (meter per day), θr 01.024; S0the Wynn and Liehr 2001).
2726 Water Air Soil Pollut (2012) 223:2719–2742

8 
2.2.4 Nitrogen Submodel >
>
DON
>
>  MicrobN  HTgrowth
>
> TON þ NH4
>
>
>
> TOC
Nitrogen basically comprises organic and inorganic >
> if > MicrobC:N
>
>
(e.g., NO3−-N, NH4+-N) nitrogen in wetland systems. < TON
DONimob ¼
To model the nitrogen cycle in FWS wetlands, the >
>  
>
> DOC  
following processes are considered: ammonification, >
>  MicrobN  HTgrowth  NH4HT
>
> TOC
imob
>
>
immobilization, nitrification, denitrification, leaching, >
>
>
> TOC
fixation, peat accumulation, and biological uptake by : if < MicrobC:N ;
TON
plants (Lee 1999; Wynn and Liehr 2001; Mayo and
ð8Þ
Bigambo 2005; Chavan and Dennett 2008; Chavan et
al. 2008). The mass balance of PON in surface and
substrate pore water consists of the PON inflow, the
NH4HT imob ¼ MicrobN  HTgrowth  NH4 =ðTON þ NH4 Þ
sloughing of dead microbial cells, the physical degra-
dation of plants to PON, PON ammonification, PON ð9Þ
immobilization by heterotrophs, PON settling from 8
>
>
TOC
> MicrobC:N or POC < 0:1
>
> 0 if
free surface water, the resuspension of PON from the >
> TON2 3
>
>    
>
>
bottom substrate layer, the accumulation of nitrogen in <  POC  6 HTgrowth  PON
> 7
peat, and PON outflow. DON in surface and substrate ¼ 46 YHT POC 7
PONammo
> TOC 5
>
>  
pore water is increased by DON inflow, the leaching >
>  Microb  HT growth  NH4HT
>
>
N imob
>
>
of N from standing dead plants, and nitrogen fixation; >
> TOC
: if < MicrobC:N ;
and it is reduced by DON ammonification, DON immo- TON

bilization by heterotrophs, plant DON uptake, and DON ð10Þ


outflow. NH4 in surface and substrate pore water com-
prises the NH4+-N inflow, nitrification of NH4+-N to 8
NO3−N, organism NH4+-N uptake, diffusion of NH4+- >
>
> 0 if
TOC
> MicrobC:N or DOC < 0:1
>
> TON
N from one pool to the other, and NH4+-N outflow. NO3 >
>
>
>
in surface and substrate pore water consists of NO3−-N <     
DONammo ¼ DOC HTgrowth DON
inflow, plant NO3−-N uptake, NO3−-N denitrification, >
>
>
   DONimob
>
>
TOC YHT DOC
diffusion of NO3−-N from one pool to the other, and >
>
>
>
:
TOC
< MicrobC:N ;
NO3−-N outflow. The mass transfer of nitrogen (DON, if
TON
NH4+-N, and NO3−-N) between free surface and sub- ð11Þ
strate pore water through diffusion was estimated in the
same way as that used for DOC. where, PONimob 0PON immobilization (gram per day),
The utilization of nitrogen by microbes through DON imob 0DON immobilization (gram per day),
immobilization and ammonification is affected by the PON ammo 0PON ammonification (gram per day),
relative fractions of carbon and nitrogen in microbes DONammo 0DON ammonification (gram per day),
and their environment (Lee 1999): NH4HT_imob 0the NH4+-H utilized by heterotrophs dur-
ing organic degradation (gram per day), and MicrobN 0
the microbial N content (grams N per gram microbes).
8  The denitrification process follows the Arrhenius
>
>
PON
 MicrobN  HTgrowth
>
> kinetics equation in the temperature range 3–28°C
>
> TON þ NH4
>
>
>
> TOC (Dawson and Murphy 1972); hence, the following
>
> if > MicrobC:N
>
> TON Arrhenius kinetics equation is used to model denitri-
<
PONimob ¼ fication (Mayo and Bigambo 2005):
>
>  
>
> POC  
>
>  MicrobN  HTgrowth  NH4HT
>
> TOC
imob
NO3denitro ¼ Dr20  θ1 ðT 20Þ  NO3 ; ð12Þ
>
>
>
>
>
> TOC
: if < MicrobC:N ;
TON
where, NO3denitro 0the denitrification of nitrate to ni-
ð7Þ trogen gas (gram per day); Dr20 0the denitrification
Water Air Soil Pollut (2012) 223:2719–2742 2727

rate constant at 20°C (one per day), θ1 0the Arrhenius attachment to accomplish the settling, resuspension, or
constant, which ranges from 1.02 to 1.09 (Tchobanoglous transformation of phosphorous. Following the works of
and Burton 1991; Beutel et al. 2009); T0the water tem- Christensen et al. (1994), Lee (1999), and Chavan and
perature (degree Celcius); and NO3 0the amount of nitrate Dennett (2008), we adjust the content of particulate
in the system (gram). phosphorous (PP) by PP inflow, PP settling, resuspen-
sion of PP, PP mineralization, and PP outflow. The
2.3 Total Suspended Sediment Submodel
settling and resuspension of particulate phosphorous,
which are directly related to sediment deposition and
In the TSS submodel, the mass balance for sediments
resuspension modeled in the sediment subsystem, are
in the wetland system, adapted from Christensen et al.
determined by the amount of sediment deposition and
(1994) and Chavan and Dennett (2008), includes the
resuspension multiplied by the ratio of the surface area
TSS inflow, sediment deposition from the free-surface
that is PP occupied and non-PP occupied. Mineraliza-
water layer to the substrate layer, resuspension of sedi-
tion of phosphorous is modeled as first-order equations
ments from the bottom of the wetland, and TSS outflow.
for both the surface and bottom categories of particles.
Recent studies have shown that for steady-state condi-
Dissolved phosphorous (DP) content is affected by DP
tions like those in constructed wetlands, where the flow
inflow, the diffusion of DP from one pool to the other,
velocity is small, settling is the major process of TSS
DP uptake by plants, and DP outflow. The mass transfer
removal (Reddy et al. 1999; Braskerud 2002b). The
of DP between free surface and substrate pore water
settling velocity of particles can be derived by Stokes
through diffusion is estimated in the same way as that
Law equation, which has been widely applied in the
used for the other dissolved nutrients mentioned above.
field (Kadlec and Knight 1996; Lee 1999; Braskerud
2002b; Chavan and Dennett 2008):
2.3.2 Autotroph Dynamics Submodel
 
Sg  1
Vsettling ¼ g  dss 
2
; ð13Þ In the autotroph dynamic submodel, the growth of
18  n
Nitrosomonas (NSgrowth) is estimated by Monod dual
where, Vsettling 0the settling velocity of sediment (meter substrate limitation kinetics combined with the effects
per second), g0the acceleration of gravity (meter per of the temperature and pH value (Wheaton et al. 1991;
square second), dss 0the diameter of a sediment particle Wynn and Liehr 2001):
(meter), Sg 0the specific gravity of the particle, and ν0
the kinematic viscosity of water (square meter per NSgrowth ¼ μNS  TempNS  pHNS  MNS ; ð15Þ
second).
The removal of suspended sediments by settling per    
day (TSSSettling; one per second) is calculated as fol- CNH4 CDO
μNS ¼ μmax NS   ;
CNH4 þ KNH4 NS CDO þ KDO NS
lows (Chavan and Dennett 2008):
ð16Þ
 
Vsettling
TSSsettling ¼ ; ð14Þ where, μNS 0the actual NS specific growth rate (one
H per day), μmax_NS 0the NS maximum growth rate (one
per day), TempNS 0the temperature factor, pHNS 0the
where, H0the mean water depth of the wetland (meter).
pH value factor, MNS 0the amount of NS (gram), CNH4
As water velocity increases, there is a point at
and CDO 0the concentrations of NH4+-N and DO, re-
which the shear stress will detach particles from the
spectively (gram per cubic meter), K NH4_NS 0NS
wetland bottom (Lee 1999). Therefore, a critical ve-
NH4+-N half saturation constant (gram per cubic me-
locity can be determined to predict the resuspension of
ter), and KDO_NS 0NS DO half saturation constant
material from the wetland bottom.
(gram per cubic meter). The expression of Monod
kinetics modifies the NS growth rate when oxygen
2.3.1 Phosphorous Submodel or ammonium is limited. The half saturation constants
represent substrate concentrations at which the growth
The phosphorous cycle is assumed that the suspended of NS reaches its half maximum rate. The amount of
sediment particles use their surface area for phosphorous NS death is modeled as a first-order reaction because
2728 Water Air Soil Pollut (2012) 223:2719–2742

there is little knowledge about the factors that control as the likelihood measure (Rankinen et al. 2006) de-
the death of microbes. fined as follows:
2.3.3 Heterotroph Dynamics Submodel
Lðθi jY Þ ¼ ð1  σ2i =σ2obs Þ; ð20Þ
The growth of heterotrophic bacteria (HTgrowth) is also
estimated in a similar way as estimating the growth of where, L(θi|Y) denotes the likelihood measure for the
Nitrosomonas. In wetlands, where both anaerobic and ith model simulation each state variable for parameter
aerobic microsites exist, bacteria can utilize these con- vector θi conditioned on a set of observations Y,
ditions for growth (Gidley 1995).
σ2i denotes the associated error variance for the ith
HTgrowth ¼ HTa growth þ HTan growth model and σ2obs denotes the observed variance for the
¼ MHT  ðaμaHT þ ð1  aÞμanHT Þ  TempHT  pHHT period under consideration.
According to the model structure, five model out-
ð17Þ puts including DO, BOD5, TN, TSS, and TP predic-
tions are influenced by parameters interactively. In
order to find an optimal parameter set with accuracy
   
CTOC CDO of each model output prediction, the model calibration
μaHT ¼ μmax aHT   ;
CTOC þ KTOC HT CDO þ KDO HT would be formulated as a multi-objective problem.
ð18Þ Subjectively assuming the importance of each model
output prediction is equitable, we integrate multiple
objectives with an equal weight into a single objective
   
CTOC KDO HT function. The objective function (NSEtotal) is defined
 
CTOC þ KTOC HT CDO þ KDO HT as follows:
μanHT ¼ μmax anHT  
CNO3
 ;
CNO3 þ KNO3 HT
NSEtotal ¼WDO NSEDO þWBOD5 NSEBOD5 þWTN NSETN
ð19Þ
þWTP NSETP þWTSS NSETSS ; ð21Þ
where, HTa_growth and HTan_growth 0the growth of aero-
bic and anaerobic heterotrophs respectively (gram per
day); MHT 0the amount of HT (gram); α0fraction of where WDO ¼ WBOD5 ¼ WTN ¼ WTSS ¼ WTP , NSE i
aerobic heterotroph (gram aerobic HT per gram HT); denotes the NSE values of target outputs for i0DO,
TempHT 0the temperature factor; pHHT 0the pH value BOD5,TN,TP,TSS, Wi denotes the weightings for the
factor; μaHT 0the actual aerobic HT specific growth rate NSE values of target output and subject to the equa-
(one per day); μanHT 0the actual anaerobic HT specific tions of FWS constructed wetland model.
growth rate (one per day); μ max_aHT 0the aerobic For the purpose of maintaining the accuracy of all
HT maximum growth rate (one per day); μmax_anHT 0 NSE values of target outputs at the same time, the stop
the anaerobic HT maximum growth rate (one/day); criterion is designed to terminate the calibration proce-
CTOC, CDO, and CNO3 0concentrations of TOC, DO, dure when each NSE value of target outputs (NSEDO,
and NO3−-N, respectively (gram per cubic meter); NSEBOD5, NSETN, NSETSS, NSETP) is greater than 0.5.
KTOC_HT 0HT TOC half saturation constant (gram per After Monte Carlo simulation, the best parameter
cubic meter), KDO_HT 0HT DO half saturation constant set was selected and used for the single-sensitivity
(gram per cubic meter); and KNO3_HT 0HT NO3−-N half analysis. In this study, single-sensitivity analysis was
saturation constant (gram per cubic meter). preceded by increasing and decreasing the one param-
eter value of the best parameter set at one time by 10%
2.4 Model Calibration, Validation, and Sensitivity to represent upper (positive) and lower (negative)
and Uncertainty Analysis bounds for each state variable (DO, etc.), respectively.
The sensitivity of the state variables to the parameter
In this study, the Nash—Sutcliffe coefficient of effi- in upper or lower bound is calculated as the ratio of the
ciency (NSE) is used as a goodness-of-fit criterion and relative absolute change in output to the change in one
Water Air Soil Pollut (2012) 223:2719–2742 2729

parameter value (Eq. 22; Chavan and Dennett 2008). 3 Results and Discussion
Then, the mean of sensitivity measure δj can be de-
fined as the sensitivity of the mean model output to a 3.1 Model Calibration
change in the parameter pj based on the mean square
sense (Eq. 23; Brun et al. 2002). Performance of the proposed model was evaluated by
comparing its output for simulations of DO, BOD5,
! TN, TSS, and TP outflow concentrations with that of
@xupper Pj @xlower Pj
Sij ¼ max i
upper ; ilower ð22Þ the measured values. Table 1 lists the calibrated pa-
@pj xi @pj xi
rameter values and those derived from the literature.
Most of the calibrated values did not deviate signifi-
cantly from those in the literature. Actually, many of
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi the parameter values were either identical to or located
1X n
in the range of values of the literature, including those
dj ¼ S2 ð23Þ
n i¼1 ij related to microbial growth, i.e., growth rates, death
rates, and half-saturation constants.
In the carbon system, an attempt was made to
where, xi denotes the output of state variable i increase the prediction accuracy of BOD5 effluent
with respect to the best parameter set; Pj denotes concentration by setting the calibrated values of bio-
the value of the parameter j in the best parameter mass carbon fraction (PTC) and BOD5 carbon fraction
set; @xupper
i and @xlower
i denote the change of output (BODC) at 0.400 and 0.725, respectively, instead of at
of state variable i with respect to parameter j in the values reported in the literature (0.470 and 0.820,
upper and lower bounds, respectively, compared to respectively). The biomass growth rate of plants
the output with respect to the best parameter set; (PTgrowth_r) was taken from the study by Polomski et
@pupper
j and @plower
j denote the change of parameter j al. (2008), indicating that the growth rates of some
in upper and lower bounds, respectively. In this aquatic plants, such as cattails, increased linearly and
study, the most important parameters are ranked correlated well with the levels of nitrogen and phos-
based on the δj values. A high δj implies that the phorous supplied. Next, an attempt was made to im-
value of parameter pj significantly influences the prove the accuracy of simulation results of dissolved
simulation results and a low one shows that the oxygen by magnifying the DO yield coefficients of
simulation result does not depend on parameter pj autotrophic and heterotrophic bacteria by 1–2 orders
(Brun et al. 2002). from the stoichiometry results of Wynn and Liehr
GLUE approach is implemented by sampling (2001). Specifically, the magnified coefficients were
from assumed probability models of uncertain 0.084 (gram microbes per gram O2) to 11.17 (gram
parameters and determining the propagation of the microbes per gram O2) for autotrophs and 0.813 to
uncertainty by model outputs. Accordingly, simula- 6.22 g microbes/g O2 for heterotrophs. Microbial res-
tions that achieve a likelihood value less than zero piration is proportional to microbial growth (Wynn
are rejected as nonbehavioral (Viola et al. 2009). and Liehr 2001); the oxygen yield rate of microbes
However, the GLUE algorithm can comprehensively implies the amount of microbial biomass growth by
evaluate the parameter likelihood function in a de- consuming one unit weight of oxygen. While there
fined parameter space, but requires intensive com- was deficient information about the microorganisms
putational time, such as using Monte Carlo in the study area, the species of autotrophs and heter-
simulations in the GLUE algorithm. To reduce this otrophs and their populations were estimated based on
computational intensity, in this study, uncertainty literature (Charley et al. 1980; Jorgensen 1983; Kadlec
analysis is preceded by the above individual sensi- 1986; Jorgensen et al. 1991; Wynn and Liehr 2001;
tivity analysis to confine the evaluation of the pa- Mayo and Bigambo 2005; Polomski et al. 2008).
rameter likelihood function to sensitive parameters. However, parameters related to microbes, e.g., mi-
Detailed discussions of the GLUE method can be crobe species, size of each population, and the fraction
found in Beven and Binley (1992), Beven and between dominant and subordinate communities,
Freer (2001), and Zak et al. (1997). heavily depend on the surviving conditions. Thus,
2730 Water Air Soil Pollut (2012) 223:2719–2742

magnified values of oxygen yields may attribute to surface water in several ways. For instance, emergent
different surroundings in which microbes survive. vegetation may reduce reaeration by blocking the
The Dr20 was increased from 1.0 (one per day) to wind and shading out algae.
3.74 (one per day) to improve the simulation of nitro- In this study, the BOD5 outflow simulation values
gen. Some of the parameters used to measure the correlated well with the measured effluent concentra-
carbon, nitrogen, and phosphorous systems could not tions in the first 2 months of the study period (i.e.,
be found in the literature (Table 1); hence, further January and February); however, a time lag occurred
experiments are warranted. The Arrhenius constant between the simulation and measured outflow values
for denitrification was 1.0485 within the range from March to mid May (Fig. 3b). The first highest peak
(1.020–1.100) of previous studies. of BOD5 outflow concentration occurred at 3.25 months,
Linear regression analysis results of calibration whereas the respective highest peak value in the simula-
(January to June 2004) suggest that this model tion occurred 1–2 weeks later. Despite the near impossi-
achieved an acceptable agreement between DO, bility of the composition of inflow wastewater remaining
BOD5, TN, TSS, and TP output, with NSE and R2 fixed all of the time, the ratio of POC to DOC might be
values over 0.6 in calibration (Table 2). Moreover, the attributed to a time lag between the simulation and
NSE values of simulated outflow concentrations measured outflow values. The TN in the calibration
against measured concentrations for DO and TSS are period (Fig. 3c) was slightly underestimated due to the
greater than 0.80. Figure 3a–e show the calibration low simulated values of NH4+-N and NO3−-N resulting
results of simulated effluent values against measured from the high nitrification and denitrification rates ini-
outflow concentrations of DO, BOD5, TN, TSS, and tially. Nitrification is the process whereby ammonium
TP. In each graph in Fig. 3, although replicating the (used by autotrophic bacteria as an electron source) is
variation trend in the measured outflow value (dotted converted into nitrate. This process is modeled as the
line), the simulated value (solid line) of the outflow quotient of Nitrosomonas growth and Nitrosomonas
concentration shows a smoother change curve. The yield (Lee 1999). The water temperatures of the study
simulated DO outflow concentration was overesti- area during the simulation period were measured from
mated from April to June in 2004 (Fig. 3a), which 16°C to 28°C, which was located in the optimal temper-
may have contributed to the estimation errors about ature range affecting Nitrosomonas growth (Wynn and
plant density and the hours of daylight. As Tainan is in Liehr 2001), the efficiency of nitrification is significantly
a tropical region where afternoon storms, with a sub- related to Nitrosomonas population. In the same way, the
sequent increase in cloud cover, are common; thus, process of denitrification is normally conducted by
overestimation of photosynthesis efficiency leads to microbes which also involved in nitrification; in addi-
higher simulated DO values. Kadlec (2005, 2008) also tion, the rate of denitrification is affected by Nitrosomo-
found that vegetation influences the oxygen supply to nas population as well. Since the Nitrosomonas
population was overestimated initially, the high efficien-
Table 2 Calibration and validation results for simulated out- cy of nitrification resulted in the simulation values for
flow values with measured outflow values TN (Mayo and Bigambo 2005).
Simulated outflow concentrations of TSS and TP
Water quality NSE R2
correlated with the measured values (Fig. 3d, e). Since
DO Calibration 0.814 0.610 the simulation processes of TSS and TP depend on
Validation 0.773 0.718 particle deposition and resuspension, physical factors
BOD5 Calibration 0.804 0.768 (e.g., water velocity and particle diameter) significantly
Validation 0.722 0.671 impacted the results. This study followed most studies of
TN Calibration 0.672 0.707 modeling wetland processes in assuming that particles
Validation 0.520 0.514 are spheres. This assumption should be discussed in the
TSS Calibration 0.808 0.826 future to increase the accuracy of the proposed model.
Validation 0.558 0.747 Microbial respiration is proportional to microbial growth
TP Calibration 0.545 0.606
(Wynn and Liehr 2001); the oxygen yield rate of
Validation 0.745 0.721
microbes implies the amount of microbial biomass
growth by consuming one unit weight of oxygen.
Water Air Soil Pollut (2012) 223:2719–2742 2731

Fig. 3 Comparison between measured outflow, simulated outflow, and inflow concentrations of a DO, b BOD5, c TN, d TSS, and e TP
from January to December 2004

3.2 Model Validation Figure 3 also shows the results of model validation
(the data collected between July and December 2004)
Linear regression analysis results of model validation for all five water quality items (i.e., DO, BOD5, TN,
(July–December 2004) suggest that the proposed and TP). The especially low value (i.e., lower than the
model achieved an acceptable agreement between the inflow concentration) of the measured outflow DO
DO, BOD5, TN, TSS, and TP output with NSE and R2 concentration in the first week of July may be attrib-
values exceeding 0.5 in the model validation step uted to the reduced reaeration caused by the vegetation
(Table 2). Moreover, the NSE values of simulated cover (Kadlec 2008) or to technical errors in the in-
outflow concentrations against measured concentra- strument; it was thus taken out as an outlier. The extent
tions for DO, BOD5, and TP are greater than 0.70. to which the model overestimates the outflow DO
2732 Water Air Soil Pollut (2012) 223:2719–2742

concentration from June to July may be attributed to microorganism processes increase the effectiveness
the overestimated plant photosynthesis rate resulting and accuracy of wetland models for modeling TN
from the raining season in Tainan (Fig. 3a). The sim- concentrations.
ulated values of BOD5 effluent were an almost perfect Concentrations of TSS were sensitive to disturban-
fit with the measured outflow values, except for the ces caused by plant harvesting and precipitation (Ou et
underestimation of data in November (Fig. 3b). Since al. 2006). The TSS simulation followed the trend of
many of the parameters related to vegetation, such as measured outflow concentrations and correlated well
the growth, death, and degradation rates of plants, with the values except for the first month in the vali-
were obtained from the literature (Wynn and Liehr dation period (i.e., July; Fig. 3d). Simulation results
2001; Polomski et al. 2008), the lower simulation for the TP effluent were good to fit the measured TP
values of BOD5 may be attributed to the different values, while exhibiting a smoother shape of high
degradation rates of plants in different climatic peak value between August and October (Fig. 3e).
regions. Microbial respiration is proportional to mi- Since the TP processes were estimated by first-order
crobial growth (Wynn and Liehr 2001); the oxygen equations, the variation trend of outflow TP concen-
yield rate of microbes implies the amount of microbial tration heavily depended on the inflow TP concentra-
biomass growth by consuming one unit weight of tion. Thus, the shape of the TP effluent variation curve
oxygen. Effluent TN concentrations were overesti- from August to October was similar to the shape of TP
mated from June to July and underestimated from July inflow concentration curve. Moreover, the underesti-
to August in the model validation results (Fig. 3c). In mated values of outflow TP concentration in the last
the proposed model, the main processes that reduce month of the study resulted from overestimation of the
TN are the uptake of ammonium and organic nitrogen plant uptake at the end of the growing season, i.e., the
by organisms (plants and microbes) as well as the effect of plant dormancy was more evident in the
nitrification of ammonium. Estimation errors of both actual situation. However, a year of simulation data
the autotrophic and heterotrophic bacteria populations may be insufficient to calibrate a phosphorus model
may contribute to the overestimation of TN because of due to storages turn over at the time scale of a couple
the influence and variability of microbes on nitrogen years and a wide band of stochastic phosphorus re-
transformation processes such as nitrification. Lower moval behavior by all wetland (Kadlec and Knight
efficiency of nitrification and nitrogen uptakes 1996).
resulted in the higher values of TN concentrations.
Additionally, among the several internal and external 3.3 Sensitivity Analysis
factors affecting the nitrogen uptake of plants and
microbes included the C/N ratios in organisms and Table 3 shows the single-sensitivity analysis results
the environment, growth rates of organisms, tempera- and ranks of parameter sensitivity with the best set of
ture variations, and pH values (Lee 1999; Wynn and model parameters after the calibration procedure using
Liehr 2001; Mayo and Bigambo 2005; Chavan and Monte Carlo simulation. A high δj means that the
Dennett 2008). A situation in which the ratio of TOC/ value of the parameter δj has an important influence
TON is less than that of MicrobC:N implies that mi- on the simulation result, while a value of zero indi-
crobial growth is C limited and the amounts of POC cates that the simulation result does not depend on the
and DOC utilized by heterotrophs are based on the parameter pj (Brun et al. 2002). The parameters with δj
relative fraction of each. However, if the ratio of TOC/ values greater than 0.1 are considered as sensitive
TON exceeds that of MicrobC:N, the environment is N model parameters (Table 3). Parameters with the sen-
limited for microbial growth. Microbes utilize the sitivity index approaching zero imply that individual
organic nutrients (DON and PON) by immobilization. parameters affected the simulations insignificantly.
They turn transform some of the nutrients into an However, some of the parameters marked with an aster-
inorganic form and release them back into the water isk in Table 1 influenced the outflow concentrations,
through mineralization or ammonification (Lee 1999; individually and jointly (Table 3). The DO concentra-
Wynn and Liehr 2001). Notably, this study did not tion was highly sensitive to the parameters related to
directly measure data of microorganism processes of microbes, including the maximum growth rates, half
the study wetland. However, measured data of saturation constants, and DO yield coefficients of both
Water Air Soil Pollut (2012) 223:2719–2742 2733

Table 3 Sensitivity analysis for model parameters

Sensitivity SDO,j (%)b c


SBOD5 ;j ð%Þb c STN,j (%)b c STSS,j (%)b c
STP,j (%)b c δj (%)d Rank
Parameter ja

dss 0 0 0.01 15.36 2.98 6.997 1


Θ1 0.99 0.45 14.91 0 0 6.686 2
BODC 1.48 8.03 0.57 0 0 3.660 3
μmax_NS 0.68 0.01 6.91 0 0 3.105 4
YNS 0.94 0.01 6.43 0 0 2.906 5
YHT_DO 3.04 0.01 0.93 0 0 1.422 6
KDO_NS 0.15 0.04 2.33 0 0 1.044 7
Dr20 0.34 0.15 2.28 0 0 1.033 8
μmax_aHT 1.7 0.59 0.43 0 0 0.827 9
KTOC_HT 1.43 0.6 0.34 0 0 0.710 10
YNS_DO 1.47 0 0.43 0 0 0.685 11
FacP/SS 0 0 0 0 1.49 0.666 12
μmax_anHT 1.02 0.42 0.26 0 0 0.507 13
YHT 0.1 1.1 0.12 0 0 0.497 14
PPmin 0 0 0 0 0.9 0.402 15
PCaccumulate_r 0.65 0.11 0.27 0 0 0.319 16
PC 0.65 0.11 0.27 0 0 0.319 17
KNH4_NS 0.25 0.04 0.53 0 0 0.263 18
KDO_HT 0.13 0.03 0.04 0 0 0.062 19
SSC 0.13 0.02 0.03 0 0 0.060 20
MicrobN 0.11 0.02 0.04 0 0 0.053 21
δNS 0.01 0 0.09 0 0 0.040 22
DONfixation_r 0 0 0.07 0 0 0.031 23
KNO3_HT 0.06 0.02 0.01 0 0 0.029 24
δHT 0.03 0.04 0 0 0 0.022 25
PTc 0 0 0 0 0 0.000 26
PTdegration_r 0 0 0 0 0 0.000 27
DOCleaching_r 0 0 0 0 0 0.000 28
PTC:N 0 0 0 0 0 0.000 29
PN 0 0 0 0 0 0.000 30
SSN 0 0 0 0 0 0.000 31
PTP 0 0 0 0 0 0.000 32
a
Definitions of the symbols are shown in Table 2
b
Percentiles which are smaller than 0.01% are shown as 0.00% in the table
c
Sij shows the largest sensitivity for change in parameter at upper or lower bounds
Sij ¼ maxð@p@x i =xi
upper ; @xi =xi Þ;
=pj @plower =p
j j j

@pupper
j represents the change of parameter j in upper bound (the prediction simulated by 10% increase or decrease of best parameter set);
@plower
j represents the change of parameter j in lower bound (the prediction simulated by 10% increase or decrease of best parameter set).
sffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P5
d
d j ¼ 15 Sij2
i¼1
2734 Water Air Soil Pollut (2012) 223:2719–2742

autotrophic and heterotrophic bacteria. The DO yield why the TP concentration is sensitive to the dss, which
coefficients of autotrophic and heterotrophic bacteria is a highly sensitive factor for TSS. Besides, outflow
largely affect DO concentrations because they influence TP concentrations are sensitive to P occupied sedi-
directly the respiration of microbes that belong to flows ment surface area (FacP/SS), which is also related to
of DO stokes in the proposed model. Some other param- particles. This finding validates the results of previous
eters related to biological and chemical oxidation of studies, which determined that particle settling is the
microbes (i.e., BODC, Dr20, θ1, YNS) also obviously major phosphorous removal mechanism (Reddy et al.
influenced the DO simulation. The DO concentration 1999; Braskerud 2002a; Chavan and Dennett 2008).
is inversely proportional to the microbial level since
oxygen is consumed increasingly by aerobic biological 3.4 Model Uncertainty Analysis
processes, such as microbial respiration, nitrification
and the decomposition of other organisms. This obser- For a situation in which calibration data are limited
vation confirms the findings of Wynn and Liehr (2001), and the input parameters are highly uncertain, the
who posited that excessive aerobic bacteria growth model output contains a large degree of uncertainty
caused the underestimation of DO concentrations. Ad- (Hassan et al. 2008). In such a case, a more quantita-
ditionally, the nitrogen contents in microbes and sedi- tive use of the calibration data and the calibration
ments were inadequately sensitive by adjusting 10% goodness-of-fit results is critical for appropriately
variation of the parameter values. quantifying the predictive uncertainty (Hassan et al.
BOD5 and TN concentrations were also sensitive to 2008). Moreover, as parameter sets rather than indi-
microbial parameters, which affected DO concentra- vidual parameter values are of primary importance, the
tions. This finding suggests that BOD5 and TN concen- interaction between parameters is simply taken into
trations are highly related to DO because, in wetlands, account in GLUE (Page et al. 2003). In this study, a
many of the degradation, transformation, and minerali- range of model responses rather than a single response
zation processes depend on microorganisms (Lee 1999; was predicted using all behavioral parameter sets.
Mayo and Bigambo 2005). Since heterotrophic bacteria Each parameter set was used to simulate the proposed
play an important role in carbon mineralization and model, in which the chosen likelihood measures were
decomposition, BOD5 is sensitive to populations of het- evaluated and the model as accepted as behavioral or
erotrophs and the amount of carbon in the BOD inflow. rejected as nonbehavioral on the basis of its goodness
In contrast, TN is highly sensitive to parameters of of fit (Page et al. 2003).
autotrophic bacteria (Nitrosomonas), Arrhenius constant Figures 4, 5, 6, 7, and 8 show the sensitivity of the
for denitrification (θ1), and the Dr20, which affect the above 18 individual sensitive parameters to DO,
efficiency of nitrification and denitrification (Dawson BDO5, TN, TSS, and TP, respectively. Each dot rep-
and Murphy 1972; Mayo and Bigambo 2005; Beutel et resents the NSE index of one parameter set from the
al. 2009). Compared to the effects autotrophs on TN, TN 5,000 Monte Carlo simulations. The best simulation
concentrations are also slightly sensitive to heterotroph tends to occur in a particular section of the parameter
parameters since heterotrophic bacteria are also respon- space, or where all behavioral simulations are restrict-
sible for ammonification and immobilization of nitrogen. ed to a smaller region of the parameter space than the
The particulate fraction of inflow ON also significantly given range (−30% to +30% of parameters in this
affects TN, while particulate and dissolved ON go study; Page et al. 2003). GLUE sensitivity analysis
through different reduction processes in the proposed results indicate that sensitive parameters included
model. BODC (fraction of C in influent BOD) in simulating
In the proposed model, microbial or plant related BOD5, θ1 (Arrhenius constant for denitrification) in
parameters did not affect the TP and TSS outflow simulating TN, Dr20 in simulating DO and TN, dss in
concentrations, except for the slight influence of the simulating TSS and TP, and FacP/SS in simulating TP.
plant biomass phosphorous fraction (PTP) on the TP These GLUE sensitivity analysis results also reflect
concentration (i.e., the δi of PTP were smaller than the feasible parameter spaces and demonstrate the
0.01% and shown as 0.00% in Table 3). The fact that relationship between model efficiency and model
the phosphorous cycle highly depends on the sus- parameters in this study. The scatter plots also
pended sediment cycle in the proposed model explains revealed that feasible values of the parameters range
Water Air Soil Pollut (2012) 223:2719–2742 2735

Fig. 4 Scatter plots of NSE values associated with sensitive parameters based on the GLUE method for simulating DO
2736 Water Air Soil Pollut (2012) 223:2719–2742

Fig. 5 Scatter plots of NSE values associated with sensitive parameters based on the GLUE method for simulating BOD5
Water Air Soil Pollut (2012) 223:2719–2742 2737

Fig. 6 Scatter plots of NSE values associated with sensitive parameters based on the GLUE method for simulating TN
2738 Water Air Soil Pollut (2012) 223:2719–2742

Fig. 7 Scatter plots of NSE values associated with sensitive parameters based on the GLUE method for simulating TSS
Water Air Soil Pollut (2012) 223:2719–2742 2739

Fig. 8 Scatter plots of NSE values associated with sensitive parameters based on the GLUE method for simulating TP
2740 Water Air Soil Pollut (2012) 223:2719–2742

over most of the calibration values. Apparently, little free surface water and substrate layers. As is well
or no sensitivity is available for the remaining param- known, the environments of different species of
eters for which good simulations are found across the microbes may vary, making it relatively easy to un-
full given range of parameter values, depending on the derestimate the importance of microbe-related factors
values of the other parameters in the set giving a and ultimately degrade the simulation accuracy of the
behavioral simulation. proposed model. Facilitating the design and manage-
ment of constructed wetlands requires improving the
simulation model by more closely examining environ-
4 Conclusions mental microorganisms in constructed wetlands. Fu-
ture studies should also examine how flow patterns
The proposed model incorporates and simulates most affect constructed wetlands. Further considerations
of the main nutrient cycles in constructed wetlands by about mechanisms of phosphorous cycle should be
using a system dynamics approach, which considers more detailed and could be integrated into the pre-
the complex interactions between nutrient metabo- sented phosphorous submodel. Finally, the effects of
lisms in wetland systems as well as the completeness flow patterns as well as the actual mixing conditions in
of the model calibration. Additionally, there are five constructed wetlands should be included in the future.
state variables of model output predictions (DO,
BOD5, TN, TSS, and TP) which could be intricately
related by the model parameters. The parameter sen- References
sitivity and uncertainty in wetland modeling are stud-
ied using the single parameter sensitivity and GLUE American Public Health Association (APHA), American Water
analyses. Many of the parameter values of the pro- Works Association, Water Environment Federation.
(1995). Standard Methods for the Examination of Water
posed model correlate with those found in the litera-
and Wastewater (19th ed.). Washington: American Public
ture, except for a few plant and microbe-related Health Association.
factors. Modeling results indicate that high R2 and Arheimer, B., & Wittgren, H. B. (2002). Modelling nitrogen
NSE values between the simulated and measured out- retention in potential wetlands at the catchment scale.
Ecological Engineering, 19(1), 63–80.
flow concentrations reveal that the proposed model Baca, R. G., & Arnett, R. C. (1976). A limnological model for
can reproduce accurately the seasonal trends of all five eutrophic lakes and impoundments, Battelle, Inc. Richland,
factors. Taken into account the interactions among Washington: Pacific Northwest Laboratories.
parameters, the GLUE method was successfully used Beutel, M. W., Newton, C. D., Brouillard, E. S., & Watts, R. J.
(2009). Nitrate removal in surface-flow constructed wetlands
to obtain the parameter uncertainty from the Monte
treating dilute agricultural runoff in the lower Yakima Basin,
Carlo simulations. The parameters of microorganisms Washington. Ecological Engineering, 35, 1538–1546.
are sensitive factors that influence DO, BOD5, and TN Beven, K. J., & Binley, A. (1992). The future of distributed
simulations; meanwhile, sediment diameter is the ma- models: model calibration and uncertainty prediction. Hy-
drological Processes, 6, 279–298.
jor factor influencing TP and TSS. The interactions Beven, K., & Freer, J. (2001). Equifinality, data assimilation,
and the relationships between cycles are evident dur- and uncertainty estimation in mechanistic modelling of
ing simulation of the proposed model. complex environmental systems using the GLUE method-
Further increasing the prediction accuracy of the ology. Journal of Hydrology, 249(1–4), 11–29.
Blasone, R. S., Madsen, H., & Rosbjerg, D. (2008). Uncertainty
proposed model requires more closely examining assessment of integrated distributed hydrological models
some parameters obtained from the literature, includ- using GLUE with Markov chain Monte Carlo sampling.
ing the composition of suspended sediments and sub- Journal of Hydrology, 353, 18–32.
strate liters, the species and populations of bacteria Braskerud, B. C. (2002a). Design consideration for increased
sedimentation in small wetlands treating agricultural run-
(autotrophs, aerobic heterotrophs, and anaerobic heter- off. Water Science and Technology, 45(9), 77–85.
trophs), and the actual growth rates of plants. Since Braskerud, B. C. (2002b). Measurement and modelling of phos-
living plants roots exude carbon compounds, i.e. an phorus retention in small constructed wetlands treating agri-
important factors affecting microbial growth, a future cultural non-point source pollution. 7th IWA Conference on
Wetland Systems for Water Pollution Control. Institute of
study should consider this component. Moreover, in Food Agricultural Sciences, University of Florida, 1, 75–85.
the proposed model, the same categories of microbial Brun, R., Kühni, M., Siegrist, H., Gujer, W., & Reichert, P.
parameters are assigned the same values in both the (2002). Practical identifiability of ASM2d parameters—
Water Air Soil Pollut (2012) 223:2719–2742 2741

systematic selection and tuning of parameter subsets. Water Treatment and Resource Recovery (pp. 393–419).
Research, 36(16), 4113–4127. Orlando, FL: Magnolia Publishing.
Charley, R. C., Hooper, D. G., & Mclee, A. G. (1980). Nitrifi- Kadlec, R. H. (2005). Vegetation effects on ammonium reduc-
cation kinetics in activated sludge at various temperatures tion in treatment wetlands. In J. Vymazal (Ed.), Natural
and dissolved oxygen concentrations. Water Research, 14, and constructed wetlands: nutrients, metals and manage-
1387–1396. ment (pp. 233–260). Leiden: Backhuys.
Chavan, P. V., & Dennett, K. E. (2008). Wetland simulation Kadlec, R. H. (2008). The effects of wetland vegetation and
model for nitrogen, phosphorus, and sediments retention in morphology on nitrogen process. Ecological Engineering,
constructed wetlands. Water, Air, and Soil Pollution, 187, 33(2), 126–141.
109–118. Kadlec, R. H., & Knight, R. L. (1996). Treatment wetlands.
Chavan, P. V., Dennet, K. E., Marchand, E. A., & Spurkland, L. Boca Raton: CRC Press.
E. (2008). Potential of constructed wetland in reducing Kadlec, R. H., Knight, R. L., Vymazal, J., Brix, H., Cooper, P.,
total nitrogen loading into the Truckee River. Wetlands & Haberl, R. (2000). Constructed wetlands for pollution
Ecology and Management, 16(3), 189–197. control — processes, performance, design and operation.
Christensen, N., Mitsch, W. J., & Jorgensen, S. E. (1994). A first IWA Scientific and Technical Report No. 8. London, UK:
generation ecosystem model of the Des Plains river exper- IWA Publishing.
imental wetlands. Ecological Engineering, 3, 495–521. Lee, E.R. (1999). Set-Wet: a wetland simulation model to opti-
Dawson, R. W., & Murphy, K. L. (1972). The temperature mize NPS pollution control. M.S. Thesis. Virginia State
dependency of biological denitrification. Water Research, University, Blacksburg, VA.
6, 71–73. Lee, E. R., Mostaghimi, S., & Wynn, T. M. (2002). A model to
Dean, S., Freer, J., Beven, K., et al. (2009). Uncertainty assess- enhance wetland design and optimize nonpoint source
ment of a process-based integrated catchment model of pollution control. Journal of the American Water Resour-
phosphorus. Stochastic Environmental Research and Risk ces Association, 38(1), 17–32.
Assessment, 23(7), 991–1010. Mantovan, P., & Todini, E. (2006). Hydrological forecasting
Diederik, P. L., Rousseaua, P. A., & Vanrolleghemb, N. D. P. uncertainty assessment: Incoherence of the GLUE method-
(2004). Model-based design of horizontal subsurface flow ology. Journal of Hydrology, 330(1–2), 368–381.
constructed treatment wetlands: a review. Water Research, Marsili-Libelli, S., & Checchi, N. (2005). Identification of dy-
38, 1484–1493. namic models for horizontal subsurface constructed wet-
Gidley, T.M. (1995). Development of a constructed subsurface lands. Ecological Modelling, 187, 201–218.
flow wetland simulation model. M.S. thesis. North Carolina Mashauri, D. A., & Kayombo, S. (2002). Application of the two
State University, Raleigh, NC coupled models for water quality management: facultative
Ghermandi, A., Bixio, D., & Thoeye, C. (2007). The role of free pond cum constructed wetland models. Physics and Chem-
water surface constructed wetlands as polishing step in istry of the Earth, 27, 773–781.
municipal wastewater reclamation and reuse. Science of Mayo, A. W., & Bigambo, T. (2005). Nitrogen transformation in
Total Environment, 380, 247–258. horizontal subsurface flow constructed wetlands I: model de-
Grosse, W. (1991). Thermoosmotic air transport in aquatic plants velopment. Physics and Chemistry of the Earth, 30, 658–667.
affecting growth activities and oxygen diffusion to wetlands Merz, S. K. (2000). Guidelines for using free water surface
soils. In D. A. Hammer (Ed.), Constructed wetlands for constructed wetlands to treat municipal sewage. Australia:
wastewater treatment (pp. 469–476). Chelsea, MI: Lewis. Queensland Department of Natural Resources.
Hafner, S. D., & Jewell, W. J. (2006). Predicting nitrogen and Ou, W. S., Lin, Y. F., Jing, S. R., & Lin, H. T. (2006). Perfor-
phosphorus removal in wetlands due to detritus accumula- mance of a constructed wetland–pond system for treatment
tion: a simple mechanistic model. Ecological Engineering, and reuse of wastewater from campus buildings. Water
27(1), 13–21. Environment Research, 78(12), 2369–2376.
Hassan, A., Bekhit, H., Chapman, J. (2008). Uncertainty assess- Page, T., Beven, K. J., Freer, J., & Jenkins, A. (2003). Investi-
ment of a stochastic groundwater flow model using GLUE gating the uncertainty in predicting responses to atmo-
analysis. Journal of Hydrology, 362(1–2), 89–109. spheric deposition using the model of acidification of
Hull, V., Parrella, L., & Falcucci, M. (2008). Modelling dis- groundwater in catchments (MAGIC) within a generalized
solved dynamics in coastal lagoons. Ecological Modelling, likelihood uncertainty estimation (GLUE) framework. Wa-
211, 468–480. ter, Air, and Soil Pollution, 142, 71–94.
Jansson, S. L., & Persson, J. (1982). Mineralization and immobi- Pastres, R., & Ciavatta, S. (2005). A comparison between the
lization of soil nitrogen. In F. J. Stevenson (Ed.), Nitrogen in uncertainties in model parameters and in forcing functions:
Agricultural Soils (pp. 229–252). Madison: American Soci- its application to a 3D water-quality model. Environmental
ety of Agronomy, Inc., Crop Science Society of America, Modelling & Software, 20(8), 981–989.
Inc., and Soil Science Society of America, Inc. Pohlert, T., Huisman, J. A., Breuer, L., & Frede, H. G. (2007).
Jorgensen, S. E. (1983). Application of ecological modelling in Integration of a detailed biogeochemical model into SWAT
environmental management, part A. New York: Elsevier. for improved nitrogen prediction––model development,
Jorgensen, S. E., Nielsen, S. N., & Jorgensen L. A. (1991). sensitivity, and GLUE analysis. Ecological Modelling,
Handbook of ecological parameters and ecotoxicology. 203, 215–228.
Amsterdam: Elsevier. Polomski, R. F., Taylor, M. D., Bielenberg, D. G., Bridges, W.
Kadlec, J. A. (1986). Nutrient dynamics in wetlands. In K. R. C., Klaine, S. J., & Whitwell, T. (2008). Differential nitro-
Reddy, & W. H. Smith (Eds.), Aquatic Plants for Water gen and phosphorous recovery by five aquatic garden
2742 Water Air Soil Pollut (2012) 223:2719–2742

species in laboratory-scale subsurface-constructed wet- Vymazal, J. (2007). Removal of nutrients in various types of
lands. HortScience, 43(3), 868–874. constructed wetlands. Science of the Total Environmental,
Rankinen, K., Karvonen, T., & Butterfield, D. (2006). An ap- 380, 48–65.
plication of the GLUE methodology for estimating the Wang, Y., Zhang, J., Kong, H., Inamori, Y., Xu, K., Inamori, R.,
parameters of the INCA-N model. Science of the Total & Kondo, T. (2009). A simulation model of nitrogen trans-
Environment, 365(1–3), 123–139. formation in reed constructed wetlands. Desalination, 235,
Reddy, K. R., Kadlec, R. H., & Gale, P. M. (1999). Phosphorus 93–101.
retention in streams and wetlands: a review. Environmental Weber, A. S., & Tchobanoglous, T. (1986). Prediction of nitri-
Science & Technology, 29(1), 83–146. fication in water hyacinth treatment systems. Journal Wa-
Reed, S. C., Middlebrooks, E. J., & Crites, R. W. (1995). ter Pollution Control Federation, 58, 376–380.
Natural systems for waste management and treatment Wheaton, F., Hochheimer, J., & Kaiser, G. E. (1991). Fixed film
(2nd ed.). NY: McGraw-Hill. nitrification filters for aquaculture. In D. E. Brune & J. R.
Sonavane, P. G., & Munavalli, G. R. (2009). Modeling nitrogen Tomasso (Eds.), Aquaculture and water quality (pp. 272–
removal in a constructed wetland treatment system. Water 303). Baton Rouge: The World Aquaculture Society.
Science & Technology, 60(2), 301–309. Wynn, T. M., & Liehr, S. K. (2001). Development of a con-
SPSS 10.0, 1989–2006, SPSS Inc., Chicago, IL. structed subsurface-flow wetland simulation model. Eco-
STELLA (1990–1997). High Performance Systems. Hanover, NH. logical Engineering, 16, 519–536.
Tchobanoglous, G., & Burton, F. L. (1991). Wastewater engi- Zak, S., Beven, K. J., & Reynolds, B. (1997). Uncertainty in the
neering: treatment, disposal, and reuse (3rd ed.). New modelling of critical loads: a practical methodology. Water,
York: McGraw-Hill. Air, and Soil Pollution, 98, 279–316.
Viola, F., Noto, L. V., Cannarozzo, M., et al. (2009). Daily Zhang, X. S., Srinivasan, R., & Bosch, D. (2009). Calibration
streamflow prediction with uncertainty in ephemeral catch- and uncertainty analysis of the SWAT model using genetic
ments using the GLUE methodology. Physics and Chem- algorithms and Bayesian model averaging. Journal of Hy-
istry of the Earth, 34(10–12), 701–706. drology, 374(3–4), 307–317.

You might also like