Analysis of neural networks as a means of developing optimal reservoir operating policies

Project investigates a neural network approach to reservoir operations.. The techniques traditionally used to generate optimal or near optimal operating policies for reservoirs fall into two general categories, optimisation and simulation. These techniques have, to date, been unable to satisfactorily address the complete range of issues which must be considered in the development of optimal reservoir operating policies. Neural networks appear to have the potential to address some of the problems associated with the use of optimisation and simulation to define the wide range of issues involved in optimal decision for operations of reservoirs. A neural network approach to reservoir operations is investigated in this thesis. In particular, the sensitivity of the performance of the neural network to variations in the key parameters of; normalisation method, number of training patterns, number of hidden neurons, learning rate, momentum coefficient, number of training iterations, and starting connection weights is examined. The approach utilises historical patterns of inflow, storage, and demand for a reservoir to predict the quantity of water to be released from the reservoir. Based on the analysis in this study, it was determined that relatively few rules exist for selection of the values of the parameters of the neural network. In general, where rules do exist, they tend to be in the form of guidelines or principles rather than explicit rules. Despite this lack of explicit rules, the neural network used in this study was able to perform reasonably well in predicting the quantity of water to be released. Overall, the concept of neural networks appear to have considerable potential to provide an alternative method for use in optimising reservoir performance, compared to the more conventional methods of optimisation or simulation.