model calibration using

30
Model calibration using

Upload: felcia

Post on 19-Jan-2016

59 views

Category:

Documents


1 download

DESCRIPTION

Model calibration using. Major steps of hydrologic modeling. (Hydrologic) data Relate to model inputs. Parameter estimation or model calibration. Model -- parameters. Predictions. Prediction uncertainty. Societal decisions. What is Model Calibration?. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Model calibration using

Model calibration using

Page 2: Model calibration using

Pag.04/21/23 2PEST program

Page 3: Model calibration using

Pag.

Major steps of hydrologic modeling

(Hydrologic) dataRelate to model inputs

Parameter estimation or model calibration

Model -- parameters

Predictions

Prediction uncertainty

Societal decisions

Page 4: Model calibration using

Pag.

What is Model Calibration?

04/21/23 4PEST program

Adjustment of the parameters of a mathematical or numerical model in order to optimize the agreement between observed data and the model's predictions.American meteorological society

The process of adjusting model inputs so that modelcalculations match what we measure in the real-world.

So, a good model calibration is not sufficient to develop a good model. Also need good data, a good model, and an adequate optimization method.

Page 5: Model calibration using

Pag.

Real world data

04/21/23 5PEST program

Page 6: Model calibration using

Pag.

Attempt 1 for model calibration

04/21/23 6PEST program

Page 7: Model calibration using

Pag.

Attempt 2 for model calibration

04/21/23 7PEST program

Page 8: Model calibration using

Pag.

Attempt 3 for model calibration

04/21/23 8PEST program

Page 9: Model calibration using

Pag.

Calculation of residuals for model calibration

04/21/23 9PEST program

Page 10: Model calibration using

Pag.

Statistical measures involving residuals

04/21/23 10PEST program

Page 11: Model calibration using

Pag.04/21/23 11PEST program

PEST is a nonlinear parameter estimation package capable of estimating parameters for any computer model. It solves a nonlinear least squares problem and minimizes the differences between the model's outputs and field measurements such as the calculated and measured discharges.

PEST adapts to the model, the model does not need to adapt to PEST.

Page 12: Model calibration using

Pag.04/21/23 12PEST program

Minmizing Sum of Squared Residuals (SSR)

The use of this equation, as an objective function to be minimized, implies certain assumptions about the residuals (Clarke, 1973):

a) That the time series of residuals have zero mean and constant variance

b) That the time series of residuals does not have a significant autocorrelation

If confidence intervals are to be given for the estimated model parameters, then:

c) residuals have to be distributed normally.

Page 13: Model calibration using

Pag.

The role of PEST

04/21/23 13PEST program

1. (initial) Parameter Sensitivity Analysis : conducted on a larger parameter space, which allows excluding from the further analysis those parameters that have a relatively small impact on the model response.

2. Calibration: (parameter estimation): PEST “calibrates” a model by reducing the discrepancies between model outputs and field observations to a minimum in the weighted least squares sense.

3. Predictive analysis: Once a parameter set has been determined for which model behavior matches system behavior as well as possible, it is then reasonable to ask whether another parameter set exists which also results in reasonable simulation by the model of the system under study.

Page 14: Model calibration using

Pag.

PEST template, instruction, and control files

04/21/23 14PEST program

1. Template file: tells PEST how to write parameters in the model input file

2. Instruction file: tells PEST where it has to look for the model-generated output, and how to read it e.g. which column contains the simulated discharges.

3. PEST conrol file: contains:User-supplied initial parameter values, range of permissible values that a parameter can take, observations, maximum number of iterations, etc.

Page 15: Model calibration using

Pag.04/21/23 15PEST program

Optimization algorithm can be classified into two categories:

1.Global search e.g. Shuffled Complex Evolution-University of Arizona (SCE-UA, 1992). searches the entire parameter space and do a controlled random search in the direction of global optimum. 2. Local search e.g. PEST program (Doherty, 1994). Estimate parameters with observations using a gradient-based method.

Gradient methods are much faster but can get stuck in local minima.

Page 16: Model calibration using

Pag.

Local and global minima

04/21/23 16PEST program

Response surface for an inverted goal function. Local minima are represented as peaks.

It shows that for any given goal function, there exits many parameter sets for which the goal functions are not significantly different from each other, i.e., there are many potential solutions based on quite different parameter sets.

Page 17: Model calibration using

Pag.04/21/23 17PEST program

Based on Objective Function, the automated calibration procedures can be classified as:

1.Single objective procedures e.g. mean squared errors in PEST program.

PEST trys to minimize the objective function in order to obtain a better fit between predicted and observed values. PEST offers the best parameter set.

2. Multiple objective procedures.

The generated output from this approach is a set of solutions (called pareto solutions) instead of a unique solution.

Page 18: Model calibration using

Pag.

When is a Model Calibrated?

04/21/23 18PEST program

As the problem of parameter optimization is not unique, it is important that we define when a model is calibrated and what the magnitude of the prediction uncertainty is.

Page 19: Model calibration using

Pag.04/21/23 19PEST program

To see when a model is calibrated:

1.Calculate the 95% prediction uncertainty

Page 20: Model calibration using

Pag.04/21/23 20PEST program

2. Normally, more than 80 percent of obsevations should be braketet by 95% confidence intervals (the upper and lower limits of the best parameter sets)

a) is not a parameter calibration problem, b)calibration can obtain smaller uncertainty distribution, c) it can be expected that some measured data can fall outside the 95PPU.

Page 21: Model calibration using

Pag.04/21/23 21PEST program

3. Upon reaching the above criteria, if there exits a significant r2 and/or Nash-Sutcliff coefficient between the best simulation and the measured data for a calibration and a test (validation) data set, then the model can be considered calibrated.

Page 22: Model calibration using

Pag.

Calibration issues

04/21/23 22PEST program

After analysing residuals to check whether they are independent,

homoscedastic and normally distributed with zero expectation;

We need to consider following questions:

1.What is the data quality and error (data uncertainty)? How does it affect

predictions?

2.What is the sensitivity of model matches to changes in model inputs? (non-

uniqueness in solutions)

3.How good is good enough, for a well calibrated model?

Page 23: Model calibration using

Pag.04/21/23 23PEST program

Thank you

Page 24: Model calibration using

Pag.04/21/23 24PEST program

Page 25: Model calibration using

Pag.04/21/23 25PEST program

WetSpa model parameter sensitivity for the Illinois River basin

Page 26: Model calibration using

Pag.04/21/23 26PEST program

Streamflow simulation using WetSpa at Illinois River basin, Oklahoma, USA

Page 27: Model calibration using

Pag.04/21/23 27PEST program

Page 28: Model calibration using

Pag.

PEST input template for WetSpa

04/21/23 28PEST program

Page 29: Model calibration using

Pag.

PEST Instruction file for WetSpa

04/21/23 29PEST program

Page 30: Model calibration using

Pag.

PEST conrol file for WetSpa

04/21/23 30PEST program