Sei sulla pagina 1di 19

Water Resour Manage (2012) 26:18271845

DOI 10.1007/s11269-012-9989-0

Optimizing Structural Best Management Practices


Using SWAT and Genetic Algorithm to Improve Water
Quality Goals

Prakash Kaini & Kim Artita & John W. Nicklow

Received: 9 August 2010 / Accepted: 13 January 2012 /


Published online: 4 February 2012
# Springer Science+Business Media B.V. 2012

Abstract A genetic algorithm (GA), an evolutionary optimization technique, is coupled


with a semi-distributed hydrologic model, Soil and Water Assessment Tool (SWAT) to find
an optimum combination of structural Best Management Practices (BMPs) that meets the
treatment goals at a watershed scale. The structural BMPs considered in the study are
detention ponds, parallel terraces, filter strips, grassed waterways, and grade stabilization
structures which are all applicable in agricultural watersheds. The decision variables in the
optimization model are the type, size, and location of BMPs which minimize the construction
cost and simultaneously reduce sediment and nutrients to target levels at the watershed outlet.
The model is demonstrated on the Silver Creek, a sub-watershed of the Lower Kaskaskia
watershed in Illinois. The model is used to compare three different sediment and nutrient
reduction cases (i.e. 20%, 40%, and, 60%) at the watershed outlet.

Keywords Genetic algorithm . Watershed modeling . Treatment goals . SWAT . Silver


Creek . Structural BMPs

1 Introduction

Quality and quantity of river flow have been major concerns in the United States and
elsewhere because of their impact on peoples daily lives and the environment. There have
been numerous programs in effect under the Clean Water Act-1972 (CWA) to improve water
quality; however, according to the US Environmental Protection Agency (USEPA), over
20,000 river segments, lakes and estuarieswhich account over 40% of assessed waters

P. Kaini (*) : K. Artita


Department of Civil and Environmental Engineering, Southern Illinois University Carbondale (SIUC),
Mail Code 6603, Carbondale, IL 62901, USA
e-mail: pkaini@golder.com

J. W. Nicklow
Department of Civil and Environmental Engineering, College of Engineering, Southern Illinois
University Carbondale (SIUC), Mail Code 6603, Carbondale, IL 62901, USA
1828 P. Kaini et al.

still do not meet water quality standards (USEPA 2008). The major causes of impairment are
nutrients, sediments, metals, and pathogens (USEPA 2008).
All point sources of pollutants in the US are required to obtain permits under the National
Pollution Discharge Elimination System (NPDES) program, administered by the USEPA
and the states. The NPDES program has been very effective in restoring water quality by
controlling pollution due to the point sources throughout the United States (USEPA 2003).
However, due to the nonpoint sources and other nontraditional sources, such as urban storm
water discharges, the water quality problems in all types of water bodies remain unsolved
(USEPA 2003). Also, land erosion significantly reduces the productivity of agricultural
lands due to loss of top soil and nutrients (Muleta and Nicklow 2001). Therefore, with the
significant achievements in controlling point sources and the growing awareness of the
effects of nonpoint sources of pollution, in 1987 the CWA was amended to address the
nonpoint pollutant sources as well (USEPA 2003). The nonpoint sources of pollutants
comprise diffuse areas like agricultural, construction, and urban areas, and other locations
not covered in point sources. Nonpoint source pollution is not subject to federal permit
requirements (USEPA 2008). Discharge from any type of conveyance system, regardless of
the source, point or nonpoint, is considered to be coming from a point source and requires
the same permit as for point sources. The major initiatives taken by federal and states
regarding nonpoint sources include the 2002 farm bill conservation provisions, the rural
clean water program, the source water protection program, the national estuary program, the
pesticides program, and costal nonpoint pollution control program (USEPA 2003). Pollution
generated from agricultural land alone has been recognized as the major source of nonpoint
source of water pollution in the United States (Srivastava et al. 2002).
Implementation of Best Management Practices (BMPs) at various locations in a
watershed has been recognized as an effective measure to reduce the impairment of
water quality (Zhen et al. 2004). BMPs are defined as structural or nonstructural
practices which are used to improve the water quantity and quality in the receiving
water body. Structural best management practices (BMPs) have been extensively used
to control runoff, sediments and nutrients, both in urban and rural or agricultural
watersheds. Structural BMPs generally applicable in agricultural watersheds include
detention and retention basins, filter strips, grade stabilization structures, grassed
waterways, and field terraces. The state-of-the-art is to design an individual BMP
structure where required on-site. However, this could lead to the implementation of
many BMPs in the watershed, and still could not assure that the treatment goals at the
watershed outlet are achieved (Emerson et al. 2005). Harrell and Ranjithan (2003)
emphasized that fewer regional basins can be implemented in place of many on-site
basins to achieve the same treatment goals when they are implemented at a watershed
scale. Muleta and Nicklow (2001) applied GA and SWAT to identify the optimal land
use patterns and tillage practices to minimize the sediment yield from the watershed.
Arabi et al. (2006) used GA and SWAT to optimize four BMPs (filter strips, parallel
terraces, grade stabilization structures, and grassed waterways) in two small water-
sheds in Indiana and found that the optimized solution was three times more cost
effective than planned strategies for the same level of protection in term of maximum
monthly sediment, phosphorus and nitrogen loads. Tuppad et al. (2010) demonstrated
the application of SWAT in modeling various BMPs at HRU, subwatershed , and
watershed levels. Mishra et al. (2007) applied SWAT to identify the areas of high
sediment yield and design structural measures. To address the watershed scale design
of BMPs, this study formulates a single objective optimal control model (OCM)
which can optimize the cost of BMPs corresponding to the different level of annual
Optimizing Structural Best Management Practices Using SWAT 1829

sediment load and nutrients (nitrogen and phosphorus) loads control (i.e. 20%. 40%,
and 60% reduction from the existing condition) at the watershed outlet. This meth-
odology can be a useful tool in devising a cost effective BMP implementation plan in
a watershed scale to meet the Total Maximum Daily Load (TMDL) criteria where
needed. The OCM employs Soil and Water Assessment Tool (SWAT), a watershed
model, and a genetic algorithm, an evolutionary optimization technique.

2 Problem Formulation

The current problem can be stated as the design of BMPs (type, size and location) at a
watershed scale that:
Minimize: Total cost of BMPs
Subject to the following constraints:
(1) Water balance in the watershed
(2) BMP size and implementation criteria constraints
(3) Land use constraints
(4) Meets sediment reduction criteria
(5) Meets nutrients reduction criteria
Mathematically, this can be expressed as
XX
Minimize : BMPC BMPCij 1
i j

P
n
Subject to : Ak  Amax
k1
Ak Ak;sb  fr
Fbw rFBw: max
2
SedLmax rSedLmax lim
NutLmax rNutLmax lim
SedLmax g1 x; u; t
NutLmax g2 x; u; t

Where, BMPC is the total cost of BMPs implemented in a watershed. BMPCij is the cost of a
j type of BMP implemented in a subbasin i. Ak is the area of the pond in subbasin k whose
value lies within the user-defined maximum limit, Amax is the maximum limit of the total
pond area assigned to the watershed, Ak,sb is the area of subbasin k, and fr is the fraction of
the subbasin area that drains to the pond. PFmax is the maximum daily flow of the simulation
period, Fbw is the filter strip width and Fbw. max is the maximum limit of the filter strip width.
SedLmax is the maximum annual sediment load and SedLmax lim is the user defined maximum
limit of annual sediment load. The final two relationships represent simulation constraints,
where g generally represents all pertinent hydrologic and hydraulic relationships, as a
function of dependent (state) and independent (decision) variables, x and u, respectively,
and time, t. BMPs are not implemented in a subbasin where the dominant land use is either
forest or water body. Grassed waterways are applied only in waterways having contributing
1830 P. Kaini et al.

drainage area less than 1,500 acres; grade stabilization structure is applied only when the
main channel slope is steeper than 1/1000; and parallel terraces are applied only where
subbasin slope is higher than 2%.
An Optimal Control Model (OCM) is developed by coupling SWAT with a genetic
algorithm (GA) to design BMPs at a watershed scale. Figure 1 illustrates the schematic of
this approach. Types, sizes, and locations of structural BMPs are considered as decision
variables, whereas average annual sediment and nutrient loads are considered as state
variables. The structural BMPs considered in the study are detention ponds, filter strips,
grade stabilization structures, parallel terraces, and grassed waterways. The OCM searches
for the least cost combination of BMPs in the watershed that meets sediment and nutrient
loads reduction criteria defined by the user.

3 Genetic Algorithm

Genetic Algorithms (GAs) are a subset of evolutionary algorithms (EAs) that mimic
biological processes to optimize an objective function (Haupt and Haupt 1998). GAs use
the principle of natural selection and evolution through generations to search for better
solutions. Developed by Holland (1975), a GA allows a population composed of many
individuals to evolve under specified selection rules to a state that minimizes/maximizes the
objective function (Goldberg 1989). Similar to other EAs, GAs do not require derivatives or
gradient information to evaluate optimal solutions. Instead GAs utilize populations (solution
points) and guided random searches to seek a better solution. GAs can be used to execute a
global as well as local search in a large search space (Srivastava et al. 2002). The three
fundamentals required for GAs to start are the representation of the parameters to be
optimized or genes, the genetic operator, and the objective function. In this study, properties
of BMPs are represented as genes and the cost of BMPs is the objective function.
GA has demonstrated its capability to optimize complex nonlinear problems in various
disciplines. Arabi et al. (2006) found a set of BMPs that can reduce the same target level of
monthly sediment, phosphorus and nitrogen load more cost effectively by optimizing BMPs
using SWAT and GA, rather than a traditional BMP management plan. Perez-Pedini et al.
(2005) combined a distributed rainfall runoff hydrologic model with a GA to optimize the
locations of infiltration-based BMPs for storm water management. Harrell and Ranjithan
(2003) presented a GA-based detention pond design (pond configurations and land use
allocations) that can achieve the desired level of pollutant removal with a lower cost than the
traditional design (individual pond design procedure) in a watershed.

Fig. 1 Schematic representation


Simulator
of optimal control model
SWAT

State Variables Decision Variables


Sediment and BMP type, size, and
nutrients location

Optimizer
Genetic Algorithm
Optimizing Structural Best Management Practices Using SWAT 1831

4 Soil and Water Assessment Tool (SWAT)

Soil and Water Assessment Tool (SWAT) (Arnold et al. 1998; Neitsch et al. 2002a), a
hydrologic model developed by the United States Department of AgricultureAgriculture
Research Service (USDA-ARS), simulates flow, sediment, and nutrients in a watershed in a
daily time step. SWAT is a physically-based and spatially-distributed continuous time
watershed model that operates on an ArcView GIS (i.e., AVSWAT) (Di Luzio et al. 2002)
platform. Arcview provides a graphical support to extract data from digital maps, and
prepares data in the format that SWAT requires as input (Abbaspour et al. 2007). Previous
application of SWAT has shown robust hydrological estimates compared to measured values
in a wide range of topographic, soil and land use watersheds. However, with respect to
climatic conditions, SWAT generally performs better in more humid watersheds than in arid
or semi-arid watersheds (Liew et al. 2007). SWAT has been used extensively for hydrological
simulations, especially in the agriculturally-dominated watersheds.
SWAT computes surface runoff, sediment, and nutrient loads with sufficient accuracy from a
wide variety of rural watersheds, especially those dominated by agriculture (Neitsch et al.
2002a). SWAT combines different state-of-the-art components of hydrological processes, and
simulates the watershed as physically and realistically as possible (Santhi et al. 2001). The
model inputs are climate, topography, soil, vegetation, and land management practices in the
watershed. The physical processes of water movement, sediment movement, plant growth, and
nutrient cycling are modeled in SWAT. The fundamental principle of hydrological simulation in
SWAT, as in any other hydrologic model, is the mass balance in a system between the incoming
and outgoing volume (Vache et al. 2002).
Distributed watershed models like SWAT are instrumental in supporting watershed
management policies aimed at reducing nonpoint source pollutions (FitzHugh and Mackay
2000). Arnold and Fohrer (2005) summarized the major applications of SWAT in the US as
its integration in BASINS (Better Assessment Science Integrating point and Nonpoint
Sources), which is being used in several states for TMDL analysis; its application in counties
along the entire US coastline, as part of the National Coastal Pollutant Discharge Inventory;
and its application to determine the impact of brush control on water supply in eight river
basins in Texas. Bracmort et al. (2006) used SWAT to study the long-term impact of
structural BMPs on water quality. SWAT is selected in this study because of its wide
application in agricultural watersheds of different sizes, locations, and climatic conditions
throughout the United States and elsewhere.
SWAT simulates flow and sediment transport continuously in a semi-distributed way by
evaluating the impact of land uses, soil types, and climatic conditions on a receiving water
body (Benaman et al. 2005). To account for the spatial heterogeneity of climate, topography,
land use and soil, SWAT delineates the watershed into subbasins based on the user-defined
critical source area (CSA). The geographic locations of subbasins in the watershed are
known and are spatially related to one another (Neitsch et al. 2002b). CSA, the minimum
area required to initialize the stream, provides the flexibility to vary the number of sub-
watersheds in a watershed (Neitsch et al. 2002a). Input information for each subbasin is
organized into the following categories: climate, hydrologic response units (HRUs), ponds/
wetlands, groundwater, and the main channel or reach draining the subbasin (Neitsch et al.
2002a). An HRU is the lumped area with threshold percentages of land use, management
and soil type in a subbasin (Arnold et al. 1998). Most of the equations in SWAT are solved
on HRU scale. Water balance in a HRU is the basis of hydrological simulation that is
provided by four storage volumes: snow (stored volume until it melts), soil profile (02 m),
shallow aquifer (220 m) and deep aquifer (>20 m).
1832 P. Kaini et al.

4.1 Demonstration Watershed and Data

The model is developed and applied to the 1,189 km2 (459 sq. miles) Silver Creek watershed
(Fig. 2), a subwatershed of Lower Kaskaskia watershed in Illinois.
The primary land uses in the watershed are cropland, grassland, and forest; however,
urbanization is expanding in the region due to the eastward expansion of Metro Link light
rail system.
Climate data for the study area, which include the daily precipitation and maximum and
minimum daily temperatures, were obtained from the National Climate Data Center
(NCDC). Other climatic data, daily solar radiation, daily wind speed and daily relative
humidity values, required by SWAT, are generated by a weather generator during the
AVSWAT simulation (Neitsch et al. 2002a). A 30-m resolution Digital Elevation Model
(DEM), and hydrologic data (daily flow) for the hydrological station USGS-05594800,
located near Freeburg (St.Clair County) in Illinois, were obtained from US Geological
Survey (USGS).
The digital soil map and the land use map were obtained from the Natural
Resources Conservation Service (NRCS). AVSWAT delineates subbasins and HRUs
by overlapping DEM map, soil map, and land use map. For simplicity, one HRU is
considered in a subbasin, and a dominated land use and soil type represent the soil
and land use type for that HRU. With the intention of creating finer subbasion
delineations, the value of CSA is considered as 400 ha, which is almost the lowest
minimum CSA for the size of the watershed considered. AVSWAT subdivides the
Silver Creek watershed into 159 subwatersheds.

Fig. 2 Location of Silver Creek


Watershed, Illinois
Optimizing Structural Best Management Practices Using SWAT 1833

4.2 Calibration of SWAT for Demonstration Watershed

SWAT is calibrated for the Silver Creek watershed using mean daily flow and daily sediment load,
observed at station USGS 05594800, located near Freeberg, Illinois. The station lies just upstream
of the Silver Creek River and Kaskaskia River confluence. SWAT is simulated for the period of
1990 through 1996, and calibration is performed for the period of 1992 through 1994. Simulation
from 1990 through 1991 is considered as a warm-up period to reduce the initial condition impacts,
whereas the simulation from 1995 through 1996 is used for verification. Daily measured sediment
data were available only for 134 days for calibration within the calibration period (1/1/1992 to 12/
31/1994) and 113 days of data were available for validation within the validation period (1/1/
1995 to 12/31/1996). Therefore, statistical evaluations of sediment calibration are based on
these limited data. This is one of the major data limitations in model calibration.
A global sensitivity analysis (SA), which combines the Latin Hypercube Sampling (LHS)
and One-factor-at-a-time sampling method (LH-OAT method) (Griensven et al. 2006) is
applied to find the most important model parameters for calibration. SA identified 14 sensitive
parameters with respect to flow and 4 sensitive parameters with respect to sediment load
generation. The sensitive parameters are listed in Tables 1 and 2. An automatic calibration
model that makes use of an evolutionary algorithm to adjust the model parameters is applied in
this study. Manual calibration of model parameters is not only time intensive but also subjective
in terms of judging the goodness of fit (Dumedah et al. 2010). A GA-based optimization model
is applied to minimize the objective function, Root Mean Square Error (RMSE) between the
observed and the simulated flow. The objective of the calibration model is to search a set of
parameter values having minimum RMSE between the mean daily observed and the simulated
output. RMSE measures the generalized standard deviation between the observed and simu-
lated flow. It is one of the most common objective functions considered in the calibration of
watershed models (Muleta et al. 2007) and is expressed as:
v
u
u1 X N  2
RMSE t  Oj  Sj 3
N 1

Table 1 SWAT parameters for


flow calibration SN Parameter for stream flow Optimal parameter value

1 CANMXP 1.2741
2 CN2PA 67.6470
3 DELAY 0.2486
4 EPCO 0.2278
5 ESCO 1.0000
6 GWQMN 76.4690
7 GW_REVAP 0.0287
8 RCHRG_DP 0.2076
9 REVAPMN 47.0391
10 SFTMP 3.5679
11 SMFMN 5.0000
12 SMFMX 5.0000
13 SMTMP 4.3102
14 TIMP 0.6637
1834 P. Kaini et al.

Table 2 SWAT parameters for


sediment calibration SN Parameter for sediment load Optimal parameter value

1 LAT_SED 0.0000
2 PRF 0.5903
3 SPCON 0.0005
4 SPEXP 1.0811

where Oj is the observed flow, Sj is the simulated flow for the day j, and N is the total number of
days. The value of RMSE ranges from zero, for perfect representation, to infinity, for the
poorest model performance.
A combination of Nash Sutcliffe Efficiency (NSE) (Nash and Sutcliffe 1970), a dimen-
sionless parameter, and Percentage Bias (PBIAS), an error index, statistics are used to
evaluate the model performance. NSE value, which varies from negative infinity to 1, with
1 indicating the perfect fit, works best when the coefficient of variation for the observed data
set is large (ASCE 1993). NSE indicates how well the plot of observed versus simulated
value fits the 1:1 line (Santhi et al. 2001), and is calculated to evaluate the model perfor-
mance. Furthermore, NSE value equal to zero indicates that the model results are no better
estimation than what would result if the average of the observed data were used (ASCE
1993). NSE is calculated as defined by Nash and Sutcliffe (1970) in Eq. 4, and PBIAS by
Eq. 5.
N 
P 2
Oj  S j
j1
NSE 1  4
N 
P 2
Oj  O
j1

2 3
P
N
O  S 100
6 j1 j j 7
6 7
PBIAS 6 7 5
4 P
N 5
Oj
j1

where NSE is the Nash-Sutcliffe efficiency, and O is the average of the observed flows.
PBIAS, deviation of the simulated data expressed as a percentage, measures the average
tendency of simulated data to be larger or smaller than their corresponding observed data
(Gupta et al. 1999; Moriasi et al. 2007). PBIAS is the percent bias.
Moriasi et al. (2007) recommended threshold values of NSE and PBIAS for model calibration.
When NSE is greater than 0.5, and PBIAS is within 25% for stream flow, 55% for sediment
and 70% for Nitrogen and Phosphorus, calibration is considered satisfactory (Moriasi et al.
2007). These criteria are considered as a basis for flow and sediment calibration in this study.
Calibration is done in two steps: flow is calibrated first and then the sediment load. In the
first step, during flow calibration, the parameters identified as impacting sediment load are
kept constant (average values of their range); and in the second step, during sediment
calibration, the parameters found from flow calibration are kept constant to search for the
optimal values of sediment parameters. Tables 1 and 2 list the calibrated values of the model
parameters for flow and sediment respectively.
Optimizing Structural Best Management Practices Using SWAT 1835

The statistics for flow calibration and verification are presented in Table 3, and the
statistics for sediment calibration and verification are presented in Table 4.
Time series plots of simulated versus measured data are compared in Figs. 3, 4, 5 and 6.
The figures show that the daily observed and simulated flows match closely; however, the
simulated peak flows are lower than the observed peak flows in almost all cases. This could
be due to one or more of the other uncertainties: errors in input, errors in observed data, or
errors in the model itself. In the case of sediment calibration and validation, the graphical
plot shows more differences in simulated and measured data. Insufficient observed sediment
data, and other uncertainties (like in the case of flow calibration) are expected to be the
causes of lower performance of sediment calibration.
The results obtained here, NSE equal to 0.7 for flow calibration and 0.83 for flow validation
based on daily data, are higher than the generally acceptable minimum NSE value (0.5) for flow
calibration (Moriasi et al. 2007). NSE values for sediment calibration and validation obtained
are 0.50 and 0.51, respectively, which are within the acceptable limits. The fact that PBIAS for
both flow and sediment calibration and validation are also within the acceptable limits justifies
the sufficient calibration and validation of SWAT for the demonstration watershed, the Silver
Creek watershed. For comparison, the coefficient of determination (R2) for calibration and
validation are evaluated and presented in Tables 3 and 4.

4.3 Representation of Structural Best Management Practices in SWAT

SWAT can incorporate different structural and non-structural BMPs simultaneously, and
simulate the watershed response. Detention ponds, filter strips, parallel terraces, grassed
waterways, and grade stabilization structures are the structural BMPs considered in this
study. A detention pond is a permanent pool that is effective in reducing peak flow as well as
trapping sediment and nutrient loads from the storm water, by retaining flow for certain time.
Sedimentation and biological processes are the principal processes that take place in the
pond to reduce suspended sediments, metals, and dissolved nutrients. It also helps to manage
the storm peaks by delaying the surface runoff before the water reaches a river. Storm water
is directed to the pond and retained in the pool until the next storm replaces it. The
performance of the detention pond depends on the size of the pond (volume), retention time
and the rainfall intensity in the watershed. Two different types of detention ponds are
considered: a normal detention pond (no infiltration) and the infiltration-type detention
pond, which has a permeable bottom represented by the coefficient of permeability (K)
equal to 1.0 mm/h. The coefficient of permeability for the normal pond is 0.0 mm/h, which
represents the impermeable bottom of the pond. Each subbasin is considered as a potential
site for a pond.
A filter strip is a uniformly-graded and densely-vegetated area at the border of the field,
where excessive sheet and rill erosion is likely to occur. Sediment, nutrient, pesticide, and
bacteria loads in surface runoff are reduced as the flow passes through the filter strip
(Neitsch et al. 2002a; Abu-Zreig et al. 2004). The vegetative strip retards the surface runoff,
and controls the reel and sheet erosion, while the sediment and pollutants are trapped, and

Table 3 Flow calibration and verification statistics

RMSE NSE R2 PBIAS

Calibration (1/1/199212/31/1994) 1.01 0.70 0.84 0.12%


Verification (1/1/199512/31/1996) 1.24 0.83 0.91 5.98%
1836 P. Kaini et al.

Table 4 Sediment calibration and verification statistics

RMSE NSE R2 PBIAS

Calibration (1/1/199212/31/1994) 624.2 0.50 0.72 25.76%


Verification (1/1/199512/31/1996) 1306.9 0.51 0.76 18.12%

also part of the flow gets infiltrated. Filter strips are planted with turf grass in most cases, but
native vegetation, such as a meadow or prairie, may also be employed. A filter strip can also
be used as the outer zone of a stream buffer. In SWAT, a filter strip is represented by a
parameter, the width of the edge of field filter strip (FILTERW). In this study, four different
widths (5, 7.5, 10, and 15 m) are considered as alternatives, whereas the default width is
zero. SWAT estimates the amount of sediment, nutrients and pesticides trapped by the filter
strip as a function of its width (Neitsch et al. 2002a), as in Eq. 6.
 0:2967
trapeff 0:367 widthfiltstrip 6

where trapeff is the trapping efficiency of the filter strip, and widthfiltstrip is the width of the
filter strip (m). Abu-Zreig et al. (2004) conducted an experiment to investigate the effect of a
vegetated filter strip on sediment removal and found the removal efficiency varying from
68% to 98% by 2 m- to 15 m-wide strips. The empirical equation used by SWAT for trapping
efficiency (Equation 4.15) gives 59% of sediment removal with a 5 m width of filter strip,
which is conservative compared to what Abu-Zreig et al. (2004) found. Schmitt et al. (1999)
also demonstrated the non-linear relationship with respect to the filter strip width. However,
considering other factors, such as vegetation type and slope of the landscape, the equation
used in SWAT is adopted in this modeling.
A grassed waterway (GWW) is a natural or constructed watercourse consisting of
vegetation and designed to accommodate concentrated flows without erosion. A grassed
waterway is capable of sustaining higher in-channel velocities than bare areas because
vegetation protects the soil by covering it and retarding flow velocity. Grassed waterways
reduce runoff velocity, filter sediment, absorb chemicals from sheet erosion, and deliver
intermittent flows to streams. GWW is represented in SWAT by modifying channel

Flow Calibration (1/1/1992-12/31/1994)


18
Sim flow
16
Obs flow
14
Flow (mm/km^2)

12

10

0
1/1/1992 7/19/1992 2/4/1993 8/23/1993 3/11/1994 9/27/1994
Time

Fig. 3 Observed and simulated daily flow in calibration


Optimizing Structural Best Management Practices Using SWAT 1837

Flow Verification (1/1/1995-12/31/1996)


35

30
Sim flow
25 Obs flow
Flow (mm/km^2)

20

15

10

0
1/1/1995 4/11/1995 7/20/1995 10/28/1995 2/5/1996 5/15/1996 8/23/1996 12/1/1996
Time

Fig. 4 Observed and simulated daily flow in verification

erodibility factor (CH_EROD), channel cover factor (CH_COV), and the Mannings roughness
coefficient (n-value) for main channel in the sub-basin (CH_N2) parameters. The default values
of CH_COV, CH_EROD both are 0.5 and CH_N2 value is 0.058 for both uplands and
lowlands. When the main channel of a sub-basin is converted to a grassed waterway, the
CH_N2 value is changed to 0.15 (Neitsch et al. 2002a) and CH_EROD and CH_COV values
are changed to 0.0. In this case, the channel segments are considered fully protected by
vegetative cover (Bracmort et al. 2006). Grassed waterways are considered in the main channel
if the contributing subbasin area is less than 1,500 hectors.
Terraces are used to reduce sheet and rill erosion and peak flow, retain moisture, and
improve water quality (USEPA 2004). Parallel terraces are formed by earth embankments or
channels, or a combination of both. Initial SCS runoff curve number for moisture condition
II (CN2), USLE support practice factor (USLE_P) and average slope length (SLSUBBSN)
are the parameters identified to represent the effects of a parallel terrace in an HRU
(Bracmort et al. 2006). CN2 is the most influential parameter for flow, sediment, and nutrient
computation (USEPA 2004). CN2 is adjusted by reducing its value by 5.0 from the existing
value, when terracing is applied. These adjusted CN2 values are close to the values

7000 Sediment Calibration (1/1/1992-12/31/1994) Sim Sediment


Obs Sediment
6000

5000
Sed (Tons)

4000

3000

2000

1000

0
1 21 41 61 81 101
Time (day)

Fig. 5 Observed and simulated sediment load in calibration


1838 P. Kaini et al.

Sediment Verification (1/1/1995-12/31/1996)


20000
18000
16000 Sim Sediment
Obs Sediment
14000
Sed (Tons)

12000
10000
8000
6000
4000
2000
0
1 21 41 61 81 101
Time (day)

Fig. 6 Observed and simulated sediment load in verification

recommended by Neitsch et al. (2002a) for terracing on different soil types, to represent
parallel terrace. When the CN2 value is reduced, the surface runoff decreases significantly,
which eventually influences the sediment and nutrient load as well. SWAT calculates the
value of sub-basin length (SLSUBBSN) from the Digital Elevation Model (DEM), and is
modified as recommended by ASAE (2003) when terracing is applied, as in Eq. 7:
100
SLSUBBSN A  S B  7
S
where SLSUBBSN is the sub-basin length, which is the horizontal spacing between terraces,
S is the average slope of the land draining into the terrace assigned by SWAT based on the
DEM, B is the variable with values 0.3, 0.6, 0.9 or 1.2, depending on soil erodibility, the
cropping system, and cropping management practices. The value of B is least for highly
erodible soils and higher values for erosion-resistant soils (ASAE 2003). A is a variable with
values from 0.12 to 0.24, depending on the location of the concerned area (ASAE 2003). A is
0.21 for this particular watershed. The USLE Support practice factor (USLE_P), with default
value 1.00, is altered to 0.4 (Neitsch et al. 2002a, b; USEPA 2004) in the HRU where a
parallel terrace is applied.
Grade stabilization structure (GSS) is a structure designed to reduce the channel
grade in natural or constructed water courses. It reduces or prevents erosion due to a
higher grade on the channel bed. Vertical drop, weir spillway, chute, and pipe drop
structures are some examples of GSS. They can be constructed using reinforced
concrete, steel sheet piling, concrete block, riprap, corrugated metal, plastic, and
concrete pipe, depending on site conditions. GSS controls head cutting or major gully

Table 5 Unit costs of BMPs


SN BMP Unit Unit Cost (US$)

1 Detention pond Acre-ft 500.0


2 Filter strip Acre 250.0
3 Parallel terrace Acre 500.0
4 Grade stabilization structure Number 6000.0
5 Grassed waterways Meter 10.0
Optimizing Structural Best Management Practices Using SWAT 1839

Table 6 Representation of BMPs in genetic algorithm

BMP Property Genes (representation in GA)

Detention pond Areaa 0 0.005 .0075 0.01 0.02


Permeabilityb 0 1.0
Contributing areac 0 0.25 0.5 0.75 0.9
Filter strip Width (m) 0 5.0 7.5 10.0 15.0
Parallel terrace Representationd 0 1
Grade S structure Slope reductione 0 0.1
Grassed waterways Mannings n 0.058 0.15
a
Fraction of subbasin area
b
Permeability (mm/h)
c
Fraction of subbasin area that contributes from to the detention pond
d
0 for no and 1 for parallel terracing application
e
Factor by which the slope of main channel is reduced

erosion in channels on steep slopes in locations where beds of intersecting channels


are at different elevations, and where flatter grades are needed in proposed channels
to control velocities.
GSSs are represented by modifying average slope of the main channel in the sub-basin
(CH_S2) and the channel erodibility factor (CH_EROD). Bracmort et al. (2006) modified
CH_S2, as in Eq. 8, by putting a structure of height D m across the waterway, and changing
CH_EROD to 0.0, the condition of reduced channel erosion, to represent GSS:

D
CH S2 CH S2SWAT ASSIGNED 8
CH L2

where, D is the height of the structure, CH_L2 is the length of the channel segment,
CH_S2SWAT_ASSIGNED is the default main channel slope, and CH_S2 is the new channel

Cost optimization by GA
60
40% redcution
20% reduction
50
60% reduction
Cost (M USD)

40

30

20

10

0
0 10 20 30 40 50 60 70 80
Generation

Fig. 7 Optimization of BMPs for three cases


1840 P. Kaini et al.

Table 7 Cost of BMPs, and sediment and nutrient loads in different cases

No reduction case 20% case 40% case 60% case

Cost (M USD) 1.036 1.494 7.461


Sediment Load (tons/year) 58,121 20,978 21,294 20,839.3
N load (kg/year) 746,585 552,586 420,760 296,247
P load (kg/year) 145,032 108,524 81,691 57,865

slope. To simplify the computation, the channel slope is reduced by 10% from its default
value where GSS is used in this study.
The unit costs of different BMPs considered in this study are presented in Table 5. These
unit costs are estimated based on personal communication at the USDANatural Resources
Conservation Service, Carbondale, IL in July 2008.
BMPs are represented as in Table 6 in the genetic algorithm for optimization.

5 Results and Discussion

An OCM was applied for three different cases: 20%, 40%, and 60% reduction on
annual sediment and nutrients (Nitrogen and Phosphorus) loads at the watershed
outlet. Figure 7 shows the progression of genetic algorithm to optimize the cost of
BMPs for three different scenarios through several generations. The maximum gener-
ation is decided based on the multiple runs beyond which there are negligible changes
in the search process (Fig. 7). Table 7 lists the costs and the sediment and nutrient
loads at the watershed outlet. Table 8 presents the number of BMPs allocated in the
watershed for three different cases.
The number of BMPs required increases as the criteria for sediment and nutrient
reduction increases. Numbers of total BMPs required for three different reduction cases,
found by the model, are 246, 171, and 148 respectively from higher to lower reduction
criteria respectively. In all cases, only one or two BMPs are assigned to most of the
subbasins. Table 9 lists the number of subbasins with different number of BMP
implementations.

Table 8 Number of BMPs for


three different cases BMP type Reduction case

20% 40% 60%

Detention pond 1 3 11
Filter strip 84 92 140
Grade stabilization structure 31 36 38
Parallel terrace 0 0 9
Grassed waterways 32 40 48
Total BMPs (no) 148 171 246
Total BMPs (1st generation) 343 324 395
Optimizing Structural Best Management Practices Using SWAT 1841

Table 9 Number of BMPs and


number of subbasins Number of BMPs Number of subbasins with BMPs in

20% case 40% case 60% case

1 70 73 66
2 32 34 56
3 4 10 20
4 0 0 2
5 0 0 0

Figures 8, 9 and 10 show the distribution of BMPs for 20%, 40% and 60% reduction
cases respectively.
Subbasins with dominant land use other than forest, wetland or any other type of
water body are qualified for BMP implementation. Out of 159 subbasins, 147 are
qualified for BMP implementation. In the first generation, a genetic algorithm assigns
different BMPs randomly to any eligible subbasin and simulates hydrological process-
es (i.e. flow, sediment and nutrients) at the watershed outlet. There are 100 solutions
in each generation and the solutions are ranked based on the cost; least cost solution
is ranked highest. OCM progresses from a random BMP selection in the initial
generation towards a more systematic BMP implementation while meeting the con-
straints or the imposed reduction criteria as well. For example, in case of 60%

Fig. 8 BMPs distribution for


20% reduction case
1842 P. Kaini et al.

Fig. 9 BMPs distribution for


40% reduction case

reduction case, the best solution found by the OCM in the first generation has 395
different BMPs assigned in different subbasins. In this case, only a single type of
BMP is implemented in 13 subbasins, 2 types of BMPs are implemented in 45
subbains, 3 types of BMPs are assigned in 63 subbasins, 4 types of BMPs are
assigned in 22 subbasins, and all 5 types of BMPs are assigned in 3 subbasins. This
is the case of the best solution among randomly generated solutions in the first
generation. However, in the final generation, for the same case, the total number of
BMPs is reduced significantly, and also only one or two BMPs are implemented as
shown in Table 9 in most of the subbasins. This pattern is similar for other sediment
and nutrient reduction criteria (i.e. 20% and 40% reduction) as well. Among the
considered BMPs, filter strip and grassed waterways are the most preferred BMPs in
all reduction cases; whereas, detention pond and parallel terrace are the least preferred
options. Parallel terraces are not considered at all in case of 20% and 40% reduction
criteria. Filter strips and grassed waterways are almost uniformly distributed through-
out the entire watershed (i.e. Figs. 6, 7 and 8). Grade stabilization structures are also
uniformly distributed; however, they are generally implemented in the outer subbasins,
which has, in general, higher channel gradient than the subbasin close the main
channel. It is remarkable to note that the ponds and parallel terraces are implemented
only when the reduction criteria is increased (i.e. 60% reduction criteria). Most of the
detention ponds selected are infiltration types (9 out of 11 detention ponds are
infiltration types). Filter strips and grassed waterways are highly effective in trapping
sediments; however, their effectiveness reduces for higher percentages reduction (i.e
Optimizing Structural Best Management Practices Using SWAT 1843

Fig. 10 BMPs distribution for


60% reduction case

60% reduction). This may be because of the dissolved portion of nutrients which go
unchecked from these two kinds of BMPs. Schmitt et al. (1999) demonstrated that
filter strips are very effective in reducing sediments but are less effective in reducing
dissolved nutrients. The reduction of dissolved nutrients is similar in case of grassed
waterways and grade stabilization structures. Therefore, for higher reduction cases,
either detention ponds or terracing options are also added. These two structures, in
fact, reduce the amount of flow generated out of the subbain.

6 Conclusion

The optimal control model, developed in this study, is able to optimize the cost of
BMPs by selecting their type, size, and locations that meets the defined level of
sediment and nutrient load reductions. The model starting from random search pro-
gresses towards implementing BMPs in strategic locations, removes less effective
BMPs, and also reduces their size so as to reduce the cost as well as meet the
treatment goals. Three different cases of reduction (20%, 40%, and 60%) were tested
and, with no surprise, the total cost of BMPs increased (i.e. number and sizes of
BMPs increase) as the sediment and nutrients reduction criteria increased. The results
showed the robust capability of the model to purposefully select the type, size, and
their locations as the reduction criteria varies. The model can be a useful tool to
implement the TMDL criteria in a cost effective manner by finding the optimal size,
type, and locations of BMPs in a watershed.
1844 P. Kaini et al.

References

Abbaspour KC, Yang J, Maximov I, Siber R, Bogner K, Mieleitner J, Zobrist J, Srinivasan R (2007) Modeling
hydrology and water quality in the pre-alpine/alpine Thur Watershed using SWAT. J Hydrol 333:413430
Abu-Zreig M, Rudra RP, Lalonde MN, Whitely HR (2004) Experimental investigation of runoff reduction and
sediment removal by vegetated filter strips. Hydrol Process 18:20292037
Arabi M, Govindaraju RS, Hantush MM (2006) Cost-effective allocation of watershed management practices
using a genetic algorithm. Water Resour Res, AGU 42:W10429. doi:10.1029/2006WR004931
Arnold JG, Fohrer R (2005) SWAT2000: current capabilities and research opportunities in applied watershed
modeling. Hydrol Process 19:563572
Arnold JG, Srinivasan R, Muttiah RS, Williams JR (1998) Large area hydrologic modeling and assessment.
Part 1: model development. J Am Water Resour Assoc 34(1):7389
ASAE (2003) Design, layout, construction and maintenance of terraces systems. ASABE, St Joseph, pp 49085
49659, 1950 Niles Road
ASCE (1993) Criteria for evaluation of watershed models. J Irrig Drain Eng 119(3):429442
Benaman J, Shoemaker C, Haith DA (2005) Calibration and validation of soil and water assessment tool on an
agricultural watershed in upstate New York. J Hydrol Eng 10(5):363374
Bracmort KS, Arabi M, Frankenberger JR, Engel BA, Arnold JG (2006) Modeling long term water quality
impact of structural BMPs. Am Soc Agric Biol Eng 49(2):367374
Di Luzio M, Srinivasan R, Arnold JG, Neitsch SL (2002) ArcView interface for SWAT2000. Blackland
Research and Extension Center, Texas Experiment Station and Grassland, Soil and Water Research Lab.,
USDA-ARS
Dumedah G, Berg AA, Wineberg M, Collier R (2010) Selecting model parameter sets from a tade-off surface
generated from the non-dominated sorting genetic algorithm. Water Resour Manag 24:44684489
Emerson CH, Welty C, Traver RG (2005) Watershed-scale evaluation of a system of storm water detention
basins. J Hydrol Eng 10(3):237242
FitzHugh TW, Mackay DS (2000) Impacts of input parameter spatial aggregation on an agricultural nonpoint
source pollution model. J Hydrol 236:3553
Goldberg DE (1989) Genetic algorithms in search, optimization and machine learning. Addison-Welsey,
Reading
Griensven A, Meixner T, Grunwald S, Bishop T, Diluzio M, Srinivasan R (2006) A global sensitivity analysis
tool for the parameters of multi-variate catchment models. J Hydrol 324:1023
Gupta HV, Sorooshian S, Yapu PO (1999) Status of automatic calibration for hydrologic models: comparison
with multilevel expert calibration. J Hydrol Eng 4(2):135143
Harrell LJ, Ranjithan SR (2003) Detention pond design and land use planning for watershed management. J
Water Resour Plan Manag 129(2):98106
Haupt RL, Haupt SE (1998) Practical genetic algorithm. Wiley, New York
Holland JH (1975) Adaptation in natural and artificial system. University of Michigan Press, Ann Arbor
Liew MWV, Veith TL, Bosch DD, Arnold JG (2007) Suitability of SWAT for the conservation effects
assessment project: comparison on USDA agricultural research service watersheds. J Hydrol Eng, ASCE
12(2):173189
Mishra A, Kar S, Singh VP (2007) Prioritizing structural management by quantifying the effect of land use
and land cover on watershed runoff and sediment yield. Water Resour Manag 21:18991913
Moriasi DN, Arnold JG, Van Liew MW, Binger RL, Harmel RD, Veith TL (2007) Model evaluation guide-
lines for systematic quantification of accuracy in watershed simulations. Am Soc Agric Biol Eng 50
(3):885900
Muleta MK, Nicklow JW (2001) Watershed management technique to control sediment yield in agriculturally
dominated areas. Water Int IWRA 26(3):435443
Muleta MK, Nicklow JW, Bekele EG (2007) Sensitivity of a distributed watershed simulation model to spatial
scale. J Hydrol Eng ASCE 12(2):163172
Nash JE, Sutcliffe JV (1970) River flow forecasting through conceptual models Part Ia discussion of
principles. J Hydrol 10:282290
Neitsch SL, Arnold JG, Kiniry, Williams JR, King KW (2002a) Soil and Water Assessment Tool, theoretical
documentation version 2000. USDA-ARS, Temple
Neitsch SL, Arnold JG, Kiniry JR, Srinivasan R, Williams JR (2002b) SWAT: Soil and Water Assessment
Tool, users manual. USDA-ARS, Temple
Perez-Pedini C, Limbrunner JF, Vogel RM (2005) Optimal locations of infiltration-based best management
practices for stormwater management. J Water Resour Plan Manag 131(6):441448
Santhi C, Arnold JG, Williams JR, Hauck LM, Dugas WA (2001) Application of watershed model to evaluate
management effects on point and nonpoint source pollution. Trans ASAE 44(6):15591570
Optimizing Structural Best Management Practices Using SWAT 1845

Schmitt TJ, Dosskey MG, Hoagland KD (1999) Filter strip performance and processes for different vegetation,
widths, and contaminants. J Environ Qual 28:14791489
Srivastava P, Hamlett JM, Robillard PD, Day RL (2002) Watershed optimization of best management
practices using AnnAGNPS and a genetic algorithm. Water Resour Res, AGU 38(3), doi:10.1029/
2001WR000365
Tuppad P, Kannan N, Srinivasan, Rossi CG, Arnold JG (2010) Simulation of agricultural management
alternatives for watershed protection. Water Resour Manag 24:31153144
USEPA (2003) National management measures to control nonpoint pollution from agriculture (http://www.
epa.gov/nps/agmm/) Accessed on: 09/20/2008, EPA-841-B-03-004
USEPA (2004) Impact of best management practices of water quality of two small watersheds in Indiana: role
of spatial scale. EPA/600/R-05/080
USEPA (2008) Overview of current total maximum daily loadTMDLprogram and regulations (http://
www.epa.gov/owow/tmdl/) Accessed on: August 27, 2008
Vache KB, Eilers JM, Santelmann MV (2002) Water quality modeling of alternative agricultural scenarios in
the U.S. Corn Belt. J Am Water Resour Assoc 383:773787
Zhen X, Yu SC, Lin J (2004) Optimal location and sizing of stormwater basins at watershed scale. J Water
Resour Plan Manag 130(4):339347

Potrebbero piacerti anche