Hybrid genetic algorithms and neural network based function approximators for water quality modeling

The principal objective of this study was to develop a computationally efficient and robust approach for the automatic calibration of surface water quality models. To accomplish this goal, an enhanced real-coded genetic algorithm (RGA) was combined with two local search methods, i.e. the Nelder-Mead simplex method (NMS) and the Powell's method (PM) of conjugate directions respectively, to form two hybrid genetic algorithms (HGAs), namely the RGA-NMS and the RGA-PM. The performances of the HGAs were compared using a CE-QUAL-W2 model and it was found that the RGA-NMS outperformed the RGA-PM in finding parameter values to minimize the objective function which was used to measure the goodness-of-fit between the observed data and simulated results. However, two problems needed to be overcome before the RGA-NMS could be put into practical use. One was the intensive computational requirement in the fitness evaluation of parameter values. The other was the fitness evaluation in the model calibration in which more than one variable was included. A neural network (NN) model was proposed to assist in overcoming the first problem. It was used to approximate the input-output response relationship underlying water quality models, and then was incorporated into the optimization algorithm to reduce the computational burden. The multi-objective model calibration technique was used to solve the second problem. Four approaches including the HGA, HGA-NN, adaptive HGA-NN and integrated HGA-NN approaches were developed for the multi-objective model calibration. These approaches used the RGA-NMS as the optimization algorithm and were different in the evaluation of the objective function. The applications of these approaches were demonstrated using a CE-QUAL-W2 model and it was found that except the HGA-NN approach, these approaches were potentially applicable to calibrating the water quality models that were of different computational requirements.