Development and Application of Artificial Neural Networks for Energy Demand Forecasting in Australia

Energy plays a very significant role in the operation of the economic machinery of a country. Insufficient energy supply can lead to high energy prices, which, in turn, can cause many, if not all, prices of commodities to increase. This leads to inflation and all its consequential adverse outcomes within the economy. For this reason, energy planning is an essential macroeconomic planning activity for the economic planning of the country. Like most other countries, energy planning in Australia is done through econo metric modelling. This is done using linear regression models that correlate energy demand (independent variable) to other macroeconomic factors, like gross domes tic product, as dependent variables. However, such a technique is unsuitable when complex interactions exist between variables. Motivated by the success of Artificial Neural Networks (ANN), this thesis aims to develop an ANN model as an alternative tool for energy demand forecasting in Australia. First, a set of macroeconomic variables is chosen as potential input features (independent variables) for the energy demand forecasting model. The output feature (dependent variable) is the monthly energy use of Australia. The output feature was measured in tons of oil. Several incremental models are developed and run to see the incremental effect of adding features on the accuracy of performance of the ANN models. Correlation analysis of features is also performed to see how each feature affects the output feature and how they are related. Then, to determine what structure of an ANN would provide better perfor mance, three different structures are chosen and run on different datasets. Dif ferent testing and training periods are used to establish the performance of each model. To automate the process of optimising ANN models, a genetic algorithm is designed to optimise the number of neurons in different ANNs. In doing this, the overall structures and performance of optimised ANNs for predicting energy production in the Australian context are presented and assessed. Also, different types of evolutionary operators are designed and tested. The results of all variants are analysed, showing the benefit of optimising the ANN structure