improving distributed hydrologocal model simulation accuracy using polynomial chaos expansion
TRANSCRIPT
M2 - Putika Ashfar Khoiri
Water Engineering LaboratoryDepartment of Civil Engineering
24th Cross-Boundary SeminarInternational Program of Maritime and Urban EngineeringOsaka University
Improving Distributed Hydrological Model Simulation Accuracy using Polynomial Chaos Expansion (PCE)
*tentative title
December 21st, 2017
1
Background of study
(Data from Japan Meteorological Agency)
There is a change in precipitation
pattern due to climate change
It is necessary to analyse rainfall-
runoff relationship to predict the
risk of flood and drought
increases due to climate change
Perform hydrological model
Input
Hydrological
Model Watershed
characteristics
Output river discharge
grid input set
2
Background of study
based on parameter complexity conceptHydrological
Model
Lumped Model
same parameter (π)in the sub-basin
Semi-distributed Model
parameters assigned in each grid cell but cells with the same parameters are grouped
π
π1
π2
π3
Fully-distributed Model
parameters assigned in each grid cell
π1
π2π3
π4
3
Background of study
based on parameter complexity conceptHydrological
Model
Fully-distributed Model
parameters assigned in each grid cell
Advantages
1. Can consider the spatial distribution of input
2. Can predict output discharge at any point
Disadvantage
1. Require many parameters so the setting and
determination of parameter is difficult
We need to assess the effectiveness of distributed
parameter including the characteristics of every
parameter
Parameter optimization is required to decrease the uncertainty
Approach:
π1
π2
π3
π4
Approach
In order to optimize the poorly known parameters and improve the model forecast
ability, data assimilation is required
Input and/or parameters
Uncertain characterisationπ₯ = [π₯1,π₯2, β¦β¦ . . π₯π]
System simulation
β’ Processβ’ Equationsβ’ Code
π¦ = π(π₯)
Outputsπ¦ = [π¦1,π¦2, β¦β¦ . . π¦π]
- Parameter optimization- Sensitivity analysis
- Distribution statistics- Performance measures
(variance, RMSE)
Characterization of uncertainty in hydrologic models is often critical for many water
resources applications (drought/ flood management, water supply utilities, reservoir
operation, sustainable water management, etc.).
4
Previous study
Previous study about parameter estimation of hydrological model in Japan
Time-dependent effect
(Tachikawa, 2014) Investigation of rainfall-runoff model in dependent flood scale
When large scale flood occurs, only roughness coefficient become a dominant
parameter because soil layer is saturated and other parameter may change over time
Spatial distributions of parameters effect
(Miyamoto, 2015) Estimation of optimum parameters sets of distributed runoff
model for multiple flood events
- Nash coefficient show the
efficiency measure of the model,
which show not good result for
medium and small floods.
5
Common method use for data assimilation :
- Variation Method : 3DVar, 4DVar- EnKF
Previous study
Improving Hydrological Model response based on the soil types, land-uses and slope classes
Example : Soil and Water Assessment Tool (SWAT) model
The turning of parameters is difficult because so many parameter need to be considered -> sensitivity analysis is needed
πππ‘ = ππ0 +
π=1
π‘
(π πππ¦ + ππ π’ππ + πΈπ β π€π πππ β πππ€)
Uncertainty analysis methods
e.g. GLUE (Generalized Likelihood Method Uncertainty Estimation) , based on Monte-Carlo simulationProblem -> Need large number of parameter sets sample
SWt = final soil water content
SW0 = initial soil water content on day i
Rday = amount of precipitation on day i
Qsurf = amount of surface runoff on day i
Ea = amount of evatranspiration on day i
Wseep = amount of water entering the vadose zone from the soil
Qgw = amount of water return flow on day i
6
Previous study
Improving Distributed Hydrological Model for the flood forecasting accuracy
Spatial distribution view Necessary to reduce grid spatial resolution (not objective)
Land-use correlated parameter
Most-considered parameter:C
Soil related parameters
-evaporation coefficient-roughness coefficient
-tank storage constant-hydraulic conductivity of layer-soil thickness-slope gradient-permeability coefficient
Therefore, I want to focus on land-use correlated parameter and soil related parameter in my study
7
Objective and Method
Polynomial Chaos Expansion (PCE)
π π₯, π‘, π =
π=0
ππππ₯
ππ(π₯, π‘) ππ(π)
πππ ππ πππ’ππ‘πππ ππ ππππ¦ππππππ
Any model output Expansion coefficient
Polynomial of order-k in parameter space determined by π
ππ π₯, π‘ =1
ππ π π₯, π‘, π ππ π π π π(π)
Density function of parameter πapproximate by GaussianQuadrature
(Mattern, 2012)
PCE has used to optimize parameter on biological ocean model
8
Objective and Method
Polynomial Chaos Expansion (PCE)
Increasing Distributed Hydrological Model Simulation Accuracy
by using Polynomial Chaos Expansion (PCE) method
Objective
The method of simulating with the value of the quadrature points in the parameter space, and estimating the optimal parameters using the difference between the observations and models (Mattern, 2012)
π πππΈγζε°
parameter space
Calculation at all quadrature points and interpolate
π πππΈ
0 1
π πππΈ
2 4 6 8 10
x 10-3
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
parameter: KDOMf
para
mete
r: r
atio
_n
0.5
1
1.5
2
2.5
3
π πππΈ
Relative parameter ranges π1
Rel
ativ
e p
aram
eter
ran
ges π
2
Global minimum of RMSE
(optimal parameter value)
Contour plot of distance function
(Hirose, 2015) 9
Method (PCE)
Emulator techniques
Polynomial Chaos Expansion (PCE)
Advantages:
- More effective than Monte-Carlo
because of their random sampling
- PCE performs a polynomial interpolation
in parameter space so it can estimate
any model output for the parameter
value of choice
Challenges:
- We must know what kind of input
parameters that make large uncertainty
- Upper and lower limits of parameter is
difficult to set
- Only two parameters can be optimized
We need to carefully consider:
1. Uncertain model input (parameters)
2. The prior distributions assigned to these
input
3. kmax (max. order of polynomial) -> we
have to check which value of kmax is
applicable for DHM
10
Previous studyImproving parameter estimation in Hydrological by applying PCE
HYMOD model (Fan, 2014) is a uniformly distributed model
Parameter Description Value
Cmax Maximum soil moisture capacity within the catchment 150-500
bexp Spatial variability of soil moisture capacity 15-5
Ξ± Distribution factor of water flowing to the quick flow reservoir 0.46
Rs Fraction of water flowing into the river from the slow flow reservoir 0.11
RqFraction of water flowing into the river from the quick flow reservoir
0.82
Soil moisture capacity function
c= soil moisture capacity
π π = 1 β 1 βπ
πͺπππ
ππππ
0 β€ π β€ πͺπππ
The PCE are applied for those two parameter, because its uniformly distributed. While another parameter is assumed to be deterministic
The results indicated both 2- and 3- order PCE's could well reflect the uncertainty of streamflow result 11
Previous study
PCE method for HYMOD model (Fan, 2014)
12
Model
Distributed Hydrological Model of Ibo River (Ishizuka, 2010)
In order to know how the impact runoff mechanism and water penetration in soil due to soil capacity can be approached by :
Storage Function method
ππ
ππ‘= π β π
π = πΎππ
π : storage heightπ : effective amount of rainfallπ: runoff heightπΎ: storage constant π: storage power constant
Runoff and slope on river drainage effect
Kinematic wave method
πβ
ππ‘+ππ
ππ₯= π
π =π sin π
πΎβ +
sin π
πβ β π·
5
3 οΌ βοΌπ·οΌ
π =sin π
πβ5
3 οΌ βοΌπ·οΌ
π : effective amount of rainfallβ : water depth π : dischargen : roughness coefficient
Try to determine couple of optimal fixed parameter that I want to optimize (e.g. π· and k in Kinematic wave method)
π : permeability coefficientπΎ : effective porosityπ· : A-layer thicknessπ : Slope gradient 13
DHM (Ishizuka, 2010)
Assumptions:
1. The storage function method in layer 1
is considered for underground
penetration after rain, because of
saturated/ unsaturated layer due to
high water depth.
2. Horizontal ground water flow is not
considered
3. Dam influence on the river way is not
considered
4. Artificial drainage system is not
considered
5. Evaporation from waterway is not
considered
6. Irrigation system is not included?
Magari
Yamazaki
Shiono
Kurisu
Tatsuno
Kamigawara
Kamae
Observation point (hourly discharge)
14
Method (Apply PCE to DHM)
Candidate for parameter optimization which are related to storage capacity of water in the soil (Ishizuka, 2010)
Runoff on slope and river channel
Description Parameter Valueslope gradient ΞΈ 0.01-13.5roughness coefficient n 0.01-2layer A thickness D 200effective porosity of layer A Ξ³ 0.2hydraulic conductivity of layer A k 0.3distance difference βx 20time interval βt 0.001storage constant of tank I K1 3.2storage constant of tank II K2 14
(Note : The range of value for each parameter should be discussed to avoid many trial and error)
15
1. Try to apply PCE to soil capacity related parameter within the catchment in
the first layer refer to Kinematic Wave equation
2γTherefore, it is necessary to check which value of polynomial order (k) is
applicable to this model
Future task
16
Test the D and k parameter with PCE to select suitable and optimum
value for the polynomial order (k) within single flood event
Determine time range for the single flood event
Study MATLAB to make modification on PCE code, apply on DHM model
Convert D and k parameter to Gaussian variables