Data treatment

EDIC-CEDEC Model

Thorsten Arnold

Aufarbeitung von GIS-Datenfür Import nach C++

Problem: - Parzellas belong to multiple Irrigation sectors

Solution: - Identify parcels belonging to Multiple sectors- Assign parcel to irrigation sector With largest relation

New problems … - Roads and rivers defined as poligons, just like plots …

Processing of Model output

Model Coupling & Sensitivity Analysis

Mod A

Mod B Mod C

Assumptions on external variables

(„World Scenario“)

PA ↔ CPA ↔ B

PB↔C

PA

PB

PC

Another model setup …Same problem for sensitivity analysis ?!?

Mod AMod B Mod C

XML - KernelData

GUI

WaSiM

Economic

Agents

XML - KernelData

GUI

CropWAT

Looking Foreward … one option

Cha

nnel

Sys

tem

WaSiM

Economic

Agents

XML - KernelData

GUI

CropWAT

Looking Foreward … another option

Channel Sys

tem

Calibration & Sensitivity

Distributed input data X

pdf :

p (X) = p (X1, X2, …, Xi)

(assumed to be known)

Realization of random variable Y

Summary Statistics:

<Y (ς) > = ∫ gς (X,P) p(X,P) dXΩ

With:

X input vectorY ouput vectorΩ k-dim space of input factorsς moment

Y= f( X,P )

Model as „blackbox“

Input Output

Parameter space & Response surface

How do changes in P affect model outputs Y ( P )?

Model results do not depend on one parameter P1. (no „turning importance“! )

Is model redundant in P1 ?(check „reducing importance“! )

Model results sensitive to both parameters P1 and P2

Sensitivity Local sensitivity in parameters Importance for calibration

Sensitivity to parameters changes with P !

P1 (-1, 1.4)

P2 (0.3,0) P3 (-0.1, -0.7)

Problem: How does my „response surface“ look like?

)(),()(

21)( P

dP

dYP

dP

dYYgrad P

Gradient

Sensitivity (V): ScreeningNumeric screening experiments

• Control experimentVary no factors: baseline run Y (P), with P = [P1, P2, …, PN]

• One-at-a-time (OAT) screeningVary one factor Pi Pi + Δ ; compare results Y (P, Pi) to control experiment Y (P)

• Factorial experimentVary all factors at the same time

(random or quasi-random representative of P from pdf, such as Latin Hypercube)

• Fractional Factorial experimentVary many factors Pi, Choose intelligent methods to save run-time

Sensitivity (V)Moris‘s OAT design

Mean ( Y (P) )

Var ( Y (P))

Increasing dynamic influence

(interaction, nonlinearity)

ΔP6

x

ΔP1

x

ΔP3

x

ΔP5

x ΔP2

x

Δ P4

x

Dynamic parameters

Linear parameters

Sensitivity & Model coupling

• How does the sensitivity of one model effect the output of other models?

• Which input data / parameter are responsible for most ouput variation / output uncertainty of each module?

• In a coupled model, how can be dealt with parameter sensitivity in order to minimize output uncertainty?