Keywords
evolutionary algorithms, meta-modelling, rainfall-runoff, artificial neural networks, calibration
Start Date
1-7-2004 12:00 AM
Abstract
Population-based search methods such as evolutionary algorithm, shuffled complex algorithm, simulated annealing and ant colony search are increasing used as automatic calibration methods for a wide range of water and environmental simulation models. However, despite the advances in computer power, it may still be impractical to rely exclusively on computationally expensive (time consuming) simulation for many real world complex problems. This paper proposed the use of meta-models to replace numerical simulation models for the purpose of calibration. Meta-models are essentially “model of the model”. The meta-model used in this study is the artificial neural network and, when coupled with genetic algorithm, forms a fast and effective hybridisation. The proposed evolutionary-based meta-model reduces the number of simulation runs required in the numerical model considerably thus making the automatic calibration of computationally intensive simulation models viable. The new approach was developed and tested in the calibration of a popular rainfall-runoff model, MIKE11/ NAM, applied to the Treggevaede catchment in Denmark. Both the calibration and verification results for single objective calibration indicate that the proposed method is able to achieve the same or better calibration performance compared to published studies using traditional population-based search methods and yet required only about 40% of the simulation runs on average.
A fast Evolutionary-based Meta-Modelling Approach for the Calibration of a Rainfall-Runoff Model
Population-based search methods such as evolutionary algorithm, shuffled complex algorithm, simulated annealing and ant colony search are increasing used as automatic calibration methods for a wide range of water and environmental simulation models. However, despite the advances in computer power, it may still be impractical to rely exclusively on computationally expensive (time consuming) simulation for many real world complex problems. This paper proposed the use of meta-models to replace numerical simulation models for the purpose of calibration. Meta-models are essentially “model of the model”. The meta-model used in this study is the artificial neural network and, when coupled with genetic algorithm, forms a fast and effective hybridisation. The proposed evolutionary-based meta-model reduces the number of simulation runs required in the numerical model considerably thus making the automatic calibration of computationally intensive simulation models viable. The new approach was developed and tested in the calibration of a popular rainfall-runoff model, MIKE11/ NAM, applied to the Treggevaede catchment in Denmark. Both the calibration and verification results for single objective calibration indicate that the proposed method is able to achieve the same or better calibration performance compared to published studies using traditional population-based search methods and yet required only about 40% of the simulation runs on average.
