Download presentation
Presentation is loading. Please wait.
1
Calibration
2
Manual Model Calibration
3
General …..
4
Non-linear parameter estimation
infer system inputs or parameters from system outputs part of the mathematics of “inverse methods” referred to as an “ill-posed problem” used in data interpretation and model calibration
5
Available Software PEST – Watermark Numerical Computing UCODE – USGS
MODFLOW2000 – USGS SCE(UA) – University of Arizona NLFIT – University of Newcastle
6
Theory …..
7
What a model does:- M o i o = M (x,p,i) Parameters p Outputs Inputs
x describes system configuration o = M (x,p,i)
8
The inverse problem:- M q i p= M-1 (x,i,q) Parameters p Measurements
Inputs i x describes system configuration p= M-1 (x,i,q)
9
Model outputs for which there are field or laboratory measurements:
excitation o1 q1 o2 q2 o3 q3 etc p1 M p2 field or laboratory measurements The model
10
The inverse problem:- M excitation q1 q2 q3 etc p1 p2 The model
field or laboratory measurements The model
11
Model domain and observation bores
12
Recharge zones
13
Contoured observations
14
Recharge zones
15
Measured and modelled water levels
16
Field or laboratory measurements:-
q1 q2 q3 x
17
Field or laboratory measurements:-
value q2 q1 q3 etc distance or time
18
Field or laboratory measurements and model output:-
p1 = p11 p2 = p21 Model output value q2 q1 q3 etc distance or time
19
Field or laboratory measurements and model output:-
p1 = p12 p2 = p22 Model output value q2 q1 q3 etc distance or time
20
Field or laboratory measurements and model output:-
p1 = p13 p2 = p23 Model output value q2 q1 q3 etc distance or time
21
Residuals:- p1 = p13 p2 = p23 value q2 q1 q3 etc distance or time
22
Residuals:- p1 = p13 p2 = p23 value q2 q1 q3 etc distance or time
23
oi ri qi
24
Objective Function:- = ri2
25
Linear system (two parameters):-
M11 p1 + M12 p2 = o1 M21 p1 + M22 p2 = o2 M31 p1 + M32 p2 = o3 M41 p1 + M42 p2 = o4 etc.
26
ie. M p = o
27
For a linear system objective function minimized when:-
p = (Mt M)-1 Mt q q1 q2 q3 q4 etc p1 p2 q = p =
28
Objective function contours
linear model p2 Objective function minimum p1
29
Objective function contours
nonlinear model p2 Objective function minimum p1
30
Objective function contours
linear model p2 Objective function minimum p1
31
Objective function contours
linear model p2 Objective function minimum p1
32
Objective function along section line
model fits data well objective function Objective function minimum parameter values
33
Objective function along section line model does not fit data well
minimum parameter values
34
Model does not fit data well or data noise is high
Model output value distance or time
35
Model does not fit data well or data noise is high
Model output based on parameter set 1 value Model output based on parameter set 2 distance or time
36
Objective function along section line model does not fit data well
Objective function is nearly at a minimum at these points Objective function minimum parameter values
37
Model fits data well or data noise is low
Model output value distance or time
38
Objective function along section line
model fits data well objective function Objective function minimum parameter values
39
Objective function contours linear model: high parameter correlation
40
Objective function along section line
linear model: high parameter correlation objective function Objective function is nearly at a minimum at these points Objective function minimum parameter values
41
Objective function contours linear model: low parameter correlation
42
Objective function along section line
linear model: low parameter correlation objective function Objective function minimum parameter values
43
Three things which make the objective function minimum less distinct:-
a large amount of data noise poor model suitability a high degree of correlation between parameters
44
It seems logical that parameters can be estimated with greater certainty in some settings than in others.
45
Stochastic Interpretation …..
46
Residuals:- value residuals q2 q1 q3 etc distance or time
47
The variance of the residuals is:- 2 = / (m - n)
m = number of observations n = number of parameters
48
The variance of the residuals is:- 2 = / (m - n)
m = number of observations n = number of parameters
49
The covariance matrix of the estimated parameter set is given by:-
C(p) = 2 (Mt M)-1
50
Covariance matrix:- 21,1 21,2 22,1 22,2 C(p1 , p2 ) = 21,2 = E{(p1 - E(p1))(p2 - E(p2))}
51
Objective function contours:-
minimum p2 p1
52
Contours of equal probability:-
Maximum probability for p1 , p2 p2 p1
53
Bivariate probability density function.
Maximum probability Bivariate probability density function.
54
Objective function contours
linear model p2 Objective function minimum p1
55
Objective function along section line
model fits data well objective function Objective function minimum parameter values
56
Objective function along section line
model fits data well maximum probability Parameter probability parameter values
57
Objective function along section line model does not fit data well
minimum parameter values
58
Objective function along section line model does not fit data well
maximum probability Parameter probability parameter values
59
Objective function contours
linear model p2 Objective function minimum p1
60
Objective function along section line model does not fit data well
minimum parameter values
61
Objective function along section line model does not fit data well
maximum probability Parameter probability parameter values
62
Objective function contours
linear model p2 Objective function minimum p1
63
Probability contour:-
64
Probability contour:-
1 1 p1
65
Probability contour:-
2 2 1 1 p1
66
Probability contour:-
67
Probability contour:-
p1+p2 p1
68
Probability contour:-
p1+p2 p1-p2 p1
69
high parameter correlation
70
insensitivity of at least one parameter
71
The covariance matrix of the estimated parameter set is given by:-
C(p) = 2 (Mt M)-1
72
The covariance matrix of the estimated parameter set is given by:-
C(p) = 2 (Mt M)-1 Increases with degree of model-to-measurement misfit Increases with degree of parameter insensitivity or correlation
73
User’s Input …..
74
Observation weights
75
Residuals:- value q2 q1 q3 etc distance or time
Assign these points a greater weight value q2 q1 q3 etc distance or time
76
Residuals:- value q2 q1 q3 etc distance or time
These points have greater weight value q2 q1 q3 etc distance or time
77
Objective Function with weights:-
= (wiri)2
78
Model domain and observation bores
79
Less data confidence
80
High predictive accuracy required here
81
High data density
82
Heads (m) very different units Conc (mg/l * 10-3)
83
Observations have large numerical range
If uniform weighting, peak flows will dominate
84
Daily Flow Water quality Monthly volume Exceedence times Φ
85
Prior Information
86
Information of the type:-
pi = x : weight = w1 a pi + b pj = y : weight = w2
87
Prior information:- p1 = a
Contribution to objective function is higher the further is p1 from a p2 a p1
88
Before prior information:-
89
After prior information:-
90
Think of prior information as:-
Additional observations and/or A penalty on the objective function Weights are important.
91
Prior information:- removes nonuniqueness
promotes stability in inversion process allows modeller to inject his/her judgement into inversion process
92
Nonlinear parameter estimation …..
93
Iterative solution improvement:-
Optimal parameters p1
94
Iterative solution improvement:-
Optimal parameters Initial parameter estimates p1
95
Iterative solution improvement:-
At least n+1 model runs per optimisation iteration (n = no. of adjustable parameters) p1
96
Repurcussions of using linearity assumption
on a nonlinear system:- convergence to objective function minimum is an iterative process calculated statistics are only approximately correct
97
More on parameter uncertainty …..
98
A Confined Aquifer head Fixed Inflow T1 T2 T3 Fixed head
99
Objective Function
100
A Confined Aquifer head Fixed Inflow T1 T2 T3 Fixed head
101
Objective Function
102
A Confined Aquifer head Fixed Inflow T1 T2 T3 Fixed head
103
A Confined Aquifer head Fixed Inflow T1 T2 T3 Fixed head
104
Hillside and Piezometers
105
System Properties Transmissivity = 100 m2/day
Creek conductance is very high Recharge = 30 mm/yr
106
Groundwater levels
107
“Observed” groundwater levels (contoured using SURFER)
108
Transmissivity distribution - I
100 m2/day
109
Transmissivity distribution - II
12 m2/day 360 m2/day
110
Irrigated area
111
Groundwater depths after irrigation - transmissivity I
0m to 1m 1m to 2m
112
Groundwater depths after irrigation - transmissivity II
0m to 1m 1m to 2m
113
SNOW section of the PERLND module of HSPF
114
PWATER section of the PERLND module of HSPF
115
PWATER section of the PERLND module of HSPF
(continued)
116
Daily Flow
117
Monthly Volume
118
Exceedence fraction
119
Parameter values lzsn 0.2732 0.8083 infilt 0.1599 0.0704
agwrc deepfr E E-03 basetp E agwetp E E-03 uzsn intfw irc lzetp
120
More parameter values lzsn 0.2732 0.8083 0.0460
infilt agwrc deepfr E E e-2 basetp E e-3 agwetp E E e-3 uzsn intfw irc lzetp
121
Bayes Theorem …..
122
p(,|yi) p(yi| ,) p(,)
123
p(,|yi) p(yi| ,) p(,)
Likelihood function of (,), ie. a measure of the degree of fitness of different parameter combinations given the actual observations, if there were no preferences for any parameter values whatsoever.
124
The posterior parameter distribution is proportional to the product of what we first thought the parameters could be, times what the data say they should be.
125
Summary …..
126
In calibrating a model we try to maximise the fit between model outcomes and field measurements
This will often allow estimation of some parameters with a high degree of certainty, and others with a low degree of certainty The more “noise” that exists in the data (and/or the less the degree to which the model is capable of simulating system fine detail), the higher will be the degree of parameter uncertainty Parameter correlation and insensitivity (which are both the same thing) increase the uncertainty associated with estimation of those parameters The more parameters that require estimation (ie. the more complex is the model) the more likely is this to occur There are quantifiable limits on what can be inferred from a given measurement dataset …..
127
The calibration process is the imposition of a set of constraints on parameter values.
That is:- When we make predictions with the model, we should only use those parameters that allow the model to match field measurement over an historical time period. This could still leave us with a wide range of parameter values to use when making model predictions.
128
Objective function minimum
p2 Optimal parameters Initial parameter estimates p1
129
Objective function minimum
“allowed parameter space” p2 p1
130
Objective function minimum
“allowed parameter space” p2 p1
131
Objective function minimum
“allowed parameter space” p2 So which parameters do we use to make a prediction? p1
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.