Open Collections

A genetic-algorithm based automatic model calibrator for the UBC watershed model

University of British Columbia

In this study, an optimization and search technique, based on the genetic algorithms (GAs) approach, is successfully used to develop an automatic model calibrator for the UBC Watershed Model. Unlike the existing random search calibration procedure, which limits the number of simultaneously calibrated modeling parameters to groups of about three to six at a time, the new GA-based calibrator allows all modeling parameters to be simultaneously evaluated. Because of the non-linear interactions between the modeling parameters, the simultaneous evaluation of all modeling parameters is demonstrated to achieve a good model calibration efficiently and quickly. The fundamental components of GAs as inspired by the Darwinian principle of natural selection are explained in detail in order to develop a complete GA-based model calibrator. A flow chart is used to illustrate the computational implementation of the GA procedures. Why GAs can work efficiently in finding an optimal set of modeling parameter values is explained by the schema theory with mathematical proofs provided. To test the soundness of the GA code developed for the automatic calibrator of the UBC Watershed Model, two well-studied watersheds in the Province of British Columbia, Campbell River and Illecillewaet River, are used. The effects of genetic operators (crossover, niching and elitism) on GA search efficiency are individually demonstrated. To objectively determine the performance of a calibrated watershed model, the difference between the observed and simulated streamflows is statistically measured. Four statistical measures are evaluated: coefficient of linear correlation (or coefficient of determination), Nash & Sutcliffe coefficient of efficiency (e!), least squares objective function and least absolute difference objective function are introduced. GA computational experiments show that the Nash & Sutcliffe coefficient of efficiency (e!) exhibits the most consistently decreasing trend of streamflow volume error (dV/V) as the coefficient value increases. A fifth statistical measure, the modified Nash & Sutcliffe coefficient (eopt!), is also used to quantify the difference between the observed and simulated streamflow data, and ensures the optimal or near-optimal set of model parameter values found at the end of a GA search achieves both high el and low dV/V at the same time.

Thesis/Dissertation

application/pdf

2009-08-06

For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use.

http://hdl.handle.net/2429/11785

Civil Engineering

University of British Columbia

UBCV

DSpace