Abstract
M.Ing. (Electrical and Electronic Engineering Science)
Scientific workflows (SWFs) and artificial neural networks (ANNs) have attracted the
attention of researchers in many fields and have been used to solve a variety of problems.
Examples of these are (a) the use of scientific workflows for the sensor web in the hydrology
domain and (b), the use of ANNs for the prediction of a number of water resource variables
such as rainfall, flow, water level and various other water quality variables. ANNs have
proved to be a powerful tool for prediction when compared with statistical methods.
The aims of this research are to develop ANNs that act as predictive models for
water resources and to deploy these models as predictive tools in a scientific workflow
environment. While there are guidelines in the literature relating to the factors affecting network
performance, there is no standard approach that is universally accepted for determining
the optimum architecture of a neural network for a given problem. The parameters of
a neural network and for the learning algorithm have a major effect on the performance
of the neural network. We consider various recurrent and feed-forward neural network
architectures for predicting changes in the water levels of dams. We explore various'
hidden layer dimensions in learning the characteristics of the training data using the
back propagation learning algorithm. Trained networks are deployed as predictive model
in a scientific workflows environment called VisTrails. ': We review and discuss the use of SWFs and ANNs in the hydrology domain with emphasis on the development of neural network architecture that will give the best predictions for water resources. A number of architectures are employed to examine the best accurate predictive network for historical rainfall data. The findings of training experiments
are promising in terms of the use of ANNs as a water resources predictive tool.
Experimental results showed how the architecture of a neural network impacts on its
predictive performance. This study shows that the number of hidden nodes is important factor for the improvement of the quality of the predictions.