#### Presentation Title

#### Keywords

river basin, decision support systems, optimization, meta-modelling, neural network, design of experiments

#### Start Date

1-7-2008 12:00 AM

#### Abstract

Evolutionary optimization algorithms, in many cases, suffer from a high computational cost due to the high-fidelity simulation models performed for objective function evaluations. Meta-modeling is one of the useful approaches to overcome this problem in which a surrogate model, running of which is much faster than the exact model, is used in lieu of the simulation model. To build a meta-model it is required to use a function-approximation procedure by which the expensive simulation model is approximated. Artificial Neural Networks (ANNs), as meta-models, have shown different applicability in various engineering design problems. However, training ANNs needs enough input-output data (design of experiments) each of which is obtained by running the expensive simulation model. A methodology is presented in this study in which the problems of design of experiments, function approximation and function optimization are sequentially solved in a feedback loop so that a much fewer number of experiments is required for the task of function approximation. The proposed approach adaptively utilizes the information obtained from function optimization, finds the regions where more data are needed, updates the training data set to fill the space and sequentially improves the accuracy of the meta-model. The performance of the proposed approach is analysed using a optimization problem on a benchmark multi-modal mathematical function and a realworld water allocation optimization problem at basin scale.

OptimumWater Allocation at Basin Scale Using Meta-Modeling

Evolutionary optimization algorithms, in many cases, suffer from a high computational cost due to the high-fidelity simulation models performed for objective function evaluations. Meta-modeling is one of the useful approaches to overcome this problem in which a surrogate model, running of which is much faster than the exact model, is used in lieu of the simulation model. To build a meta-model it is required to use a function-approximation procedure by which the expensive simulation model is approximated. Artificial Neural Networks (ANNs), as meta-models, have shown different applicability in various engineering design problems. However, training ANNs needs enough input-output data (design of experiments) each of which is obtained by running the expensive simulation model. A methodology is presented in this study in which the problems of design of experiments, function approximation and function optimization are sequentially solved in a feedback loop so that a much fewer number of experiments is required for the task of function approximation. The proposed approach adaptively utilizes the information obtained from function optimization, finds the regions where more data are needed, updates the training data set to fill the space and sequentially improves the accuracy of the meta-model. The performance of the proposed approach is analysed using a optimization problem on a benchmark multi-modal mathematical function and a realworld water allocation optimization problem at basin scale.