AUTOCAL TRAINING MODULE FOR IOWA

Report
Applying and Interpreting the SWAT
Sensitivity Analysis and
Auto-calibration Tools
by
Mike Van Liew
Dept. of Biological Systems Engineering
University of Nebraska
Lincoln, NE
Heartland Regional Water Coordination Initiative
Available Auto-calibration
tools in SWAT
--Auto-calibration tools created by Ann van Griensven
(2005)
--Tools include:
Sensitivity Analysis
Parasol
model calibration
parameter uncertainty
Sunglasses
parameter uncertainty for calibration and
and validation periods
Limitations of the ArcSWAT Interface
Auto-calibration Tool
• The ArcSWAT Interface Sensitivity Analysis/Auto-Calibration
and Uncertainty Tools only allow calibration at a single point
within a watershed
• In some cases, a multi-point, regional approach to
calibration is highly desirable, especially for large
watersheds
Running the Sensitivity
Analysis/Auto-calibration Tool in the
Project Directory
• The ArcSWAT Interface provides a framework for
constructing files that are necessary for performing
sensitivity analysis or a multi-gage, multi-parameter
calibration
• Some files employed in the Interface tools must be modified
by hand to perform a multi-gage or multi-parameter
calibration
• This can be accomplished by working in the project directory
instead of the ArcSWAT Interface
Today’s Objectives:
• Learn how to create and modify the necessary files for
running the sensitivity analysis and auto-calibration tools in
a project directory for multi-gage, multi-constituent
configurations
• Learn how to interpret the output files generated from the
sensitivity analysis and auto-calibration tools
Parameter Sensitivity
• Challenge of determining which parameters to calibrate so
that the model response mimics the actual field, subsurface,
and channel conditions as closely as possible
• Calibration process becomes complex and computationally
extensive when the number of parameters in a model is
substantial
• Sensitivity analysis can be helpful to identify and rank
parameters that have significant impact on specific model
outputs of interest
Sensitivity Analysis in SWAT
• helpful to model users in identifying parameters that are most
influential in governing streamflow or water quality response
• allows model users to conduct two types of analyses:
--the first analysis may help to identify parameters that improve a
particular process or characteristic of the model (assesses the
impact of adjusting a parameter value on some measure of
simulated output, such as average streamflow)
--second type of analysis uses measured data to provide an overall
“goodness of fit” estimation between the modeled and the
measured time series (identifies the parameters that are affected
by the characteristics of the study watershed and those to which
the given project is most sensitive)
Sensitivity Analysis
• Sensitivity analysis demonstrates the impact that change to an
individual input parameter has on the model response
• Method in SWAT combines the Latin Hypercube (LH) and Onefactor-At-a-Time (OAT) sampling
• LH = generates a distribution of plausible collections of parameter
values from a multidimensional distribution
• During sensitivity analysis, SWAT runs (p+1)*m times, where p is
the number of parameters being evaluated and m is the number
of LH intervals or loops
• For each loop, a set of parameter values is selected such that a
unique area of the parameter space is sampled
Sensitivity Analysis
• That set of parameter values is used to run a baseline
simulation for that unique area
• Then, using one-at-a-time (OAT) sampling, a parameter is
randomly selected, and its value is changed from the
previous simulation by a user-defined percentage
• SWAT is run on the new parameter set, and then a different
parameter is randomly selected and varied
• After all the parameters have been varied, the LH algorithm
locates a new sampling area by changing all the parameters
Getting Started: Building Files to
Conduct Sensitivity Analysis
• ArcSWAT Interface Sensitivity Analysis Tool Input and Output
Windows
• Manually modify files in project directory that are written
from the Interface
Sensitivity Input Window
Analysis Location: Select
from the SWAT simulation
list a simulation for
performing the sensitivity
analysis
Subbasin: Select a
subbasin within the
project where observed
data will be compared
against simulated output
Sensitivity Input Window
Hypercube intervals
(Alpha_Bf): 10 intervals of
0-0.1, 0.1-0.2 … 0.9-10.0
OAT change (Alpha_Bf):
Changes by 5% x (1.0 0.0) = 0.05
Initial value of 0.13
becomes 0.08 or 0.18
Sensitivity Input Window
Observed Data File Name
Select Parameters for
conducting sensitivity
analysis
Lower bound = 0.0
Upper bound = 10.0
Adjust if necessary
Variation Method:
1) Replace by value
2) Add to value
3) Multiply by value (%)
Sensitivity Analysis Output Window
Output Evaluation:
Comparison variable(s)
Select Average Modeled
Output (eg, streamflow)
Or Percent of Time
output is < a threshold
value)
Select Concentrations or
Loads for Water Quality
Objective Function:
Select optimization
method
Write Input Files to
Project Directory
Main Output: Sensout.out
Input Data:
Objective and
Response
Functions
List of Parameters
Sample of Senspar.out file OAT = .05 Loops = 5
run
ALPHA_BF
ESCO
CH_K2
SOL_AWC
GW_DELAY
0-1.0
0-1.0
0-150.
+ 20%
0-30
1
0.16
0.59
25.61
-0.99
8.76
2
0.16
0.59
25.61
-0.99
7.26
3
0.21
0.59
25.61
-0.99
7.26
4
0.21
0.59
25.61
-2.99
7.26
5
0.21
0.59
18.11
-2.99
7.26
6
0.21
0.64
18.11
-2.99
7.26
7
0.94
0.33
32.36
-13.83
19.30
8
0.94
0.33
39.86
-13.83
19.30
9
0.94
0.33
39.86
-13.83
20.80
10
0.99
0.33
39.86
-13.83
20.80
11
0.99
0.33
39.86
-11.83
20.80
12
0.99
0.38
39.86
-11.83
20.80
Main Output: Sensout.out
Parameter Ranking
Ranking of 16 Parameters for Mahantango Creek
Watershed, PA
Ranking of 16 Parameters for Stevens Creek
Watershed, PA
Mean Value Percent Difference in Objective Function Value
with a 5% Change in Parameter Value for Stevens Creek
Watershed, NE
Strengths of the Automated Approach
to Calibration in SWAT
• Manual calibration of a dozen or more parameters that
govern streamflow can be a very time consuming and
frustrating process
• The auto-calibration procedure in SWAT provides a powerful,
labor-saving tool that can be used to substantially reduce
the frustration and uncertainty often associated with
manual calibration
• The Parasol with Uncertainty Analysis tool in SWAT provides
optimal parameter values that are determined through an
optimization search. It also provides an indication of how
sensitive a parameter is to being precisely calibrated, based
upon the user supplied input range
Shuffled Complex Evolution
Algorithm (SCE-UA)
• calibration procedure based on a Shuffled Complex
Evolution Algorithm (SCE-UA) and a single objective function
• In a first step, the SCE-UA selects an initial population of
parameters by random sampling throughout the feasible
parameter space for “p” parameters to be optimized, based
on given parameter ranges
• The population is partitioned into several communities
(complexes), each consisting of “2p+1” points
Shuffled Complex Evolution
Algorithm (SCE-UA)
• Each community is made to evolve based on a statistical
“reproduction process” that uses the simplex method, an
algorithm that evaluates the objective function in a
systematic way with regard to the progress of the search in
previous iterations
• At periodic stages in the evolution, the entire population is
shuffled and points are reassigned to communities to ensure
information sharing
• As the search progresses, the entire population tends to
converge toward the neighborhood of global optimization,
provided the initial population size is sufficiently large
Shuffled Complex Evolution Algorithm (SCE-UA)
Select
Parents
Initialize
Shuffle
Evolve
No
Assess
Yes
End
Repeat y
times to
generate y
offspring
Generate
Offspring
Replace
Parents by
Offspring
Repeat x
times to
generate x
offspring
Limitations of the ArcSWAT Interface
Auto-calibration Tool
• The ArcSWAT Interface Sensitivity Analysis/Auto-Calibration
and Uncertainty Tools only allow calibration at a single point
within a watershed
•
In some cases, a multi-point, regional approach to
calibration is highly desirable, especially for large
watersheds
Building Files to Conduct
Auto-calibration
• ArcSWAT Interface Auto-calibration Tool Input and Output
Windows
• Manually modify files in project directory that are written
from the Interface
Auto-calibration Input Window
Analysis Location: Select
from the SWAT simulation
list a simulation for
performing the calibration
Subbasin: Select a
subbasin within the
project where observed
data will be compared
against simulated output
Auto-calibration Input Window
Optimization Settings
MAXN = Maximum number
of trials before optimization is
terminated
NGS = Number of complexes
IPROB = sets the threshold
for ParaSol:
1 = 90% CI
2 = 95% CI
3 = 97.5% CI
Observed Data File Name
Calibration Method:
ParaSol or ParaSol
with Uncertainty
Analysis
Auto-calibration Input Window:
Observed Daily Record for Streamflow
Year of
observed
record
Julien day
of
observed
record
Observed
Daily
Streamflow
in cms
Input Window: Observed Monthly Record for
Streamflow and Sediment
Year of
observed
record
Month of
observed
record
Observed
Monthly
Streamflow
(cms)
Observed
Monthly
Sediment
Load
(tons/day)
Auto-calibration Input Window
Select Parameters for
calibration
Adjust initial lower and
upper bounds, if necessary
(note: minimum lower bound
for SURLAG = 0.5)
Auto-calibration input
files:
Changepar
Parasolin
MAXN
NGS
IPROB
Auto-calibration input: Multigage Changepar file is
created by combining two or more changepar files that are
specific for certain subbasins or HRUs in the project
Upper Gage
Lower Gage
Auto-calibration input: Multigage Changepar file is
created by combining two or more changepar files that are
specific for certain subbasins or HRUs in the project
For parameters that vary by
HRU, select All Land Uses, Soils,
and Slopes for Subbasins that
are relevant to a particular gage
For parameters that vary by
Subbasin, select All Subbasins
that are relevant to a particular
gage
Auto-calibration input: Multigage Changepar file is
created by combining two or more changepar files that are
specific for certain subbasins or HRUs in the project
Subbasins
for gage 1
HRUs for
gage 1
Subbasins
for gage 2
HRUs for
gage2
Auto-calibration Output Window
Output Evaluation: Select
parameter to be calibrated
Objective Function:
Select optimization
method
Select Concentrations or
Loads for Water Quality
Calibration
Write Input Files to
Project Directory
Auto-calibration input file: fig
Autocal Command Code
and Observed Data Files
for 2 Gage Locations
Auto-calibration input file: Filecio
Number of years
simulated
ICLB =AutoCalibration
Default = 0
Sensitivity = 1
Optimization = 2
Optimization with
uncertainty = 3
Bestpar = 4
NYSKIP = Warm-up
Auto-calibration input file: Objmet
Code number for
calibration variable
Concentration or load
Given weight for
objective function
Objective function
method
Code number for
Autocalfile in .fig
Auto-calibration output file: Parasolout
Calibration
Parameter
Uncertainty
Ranges
Auto-calibration output file: goodpar
and bestpar
Parameter listings
Calibration
Auto-calibration output file: Autocal
Monthly Sediment Load
Monthly
Streamflow
Calibration
Parameter
Uncertainty
Ranges
Measured versus Simulated Streamflow with
Parasol Uncertainty CI

similar documents