AbstractThis thesis documents the research and analysis carried out in order to implement a system for short-term load forecasting for the isolated power system of the island of Crete, Greece.
The system was based on the use of multilayer perceptron neural networks. Data used to train and test the networks were obtained from the Public Power Corporation and span four years. These data were preprocessed using a special data preprocessing approach that is also introduced and described in this thesis. The data corresponding to the first year (1994) were used for the network training whereas the data for the other three years were used for testing. Extensive studies on the importance of various factors such as temperature, season, day of the week, etc. to the load demand were performed and the conclusions drawn lead to a better understanding of the load demand curve.
Various network topologies were validated so that their effect on the results could be evaluated and the best one chosen. This was done by studying the effect of factors such as learning rate, momentum, number of the training patterns etc. Also a new neural network output representation was utilized based on the use of Gray code, which provides a better error tolerance. The results show that the forecasted load average error achieved is extremely satisfactory and furthermore the majority of the erroneous predictions lie in the next output range (higher or lower).
The system is to be used with real-world data so as to provide the Public Power Corporation of Crete with the ability to forecast the load demand through the year.
|Date of Award||Oct 2000|
|Supervisor||Andrew Ware (Supervisor)|