In my Bsc thesis the methods of short-term load forecasting and the realisation of some methods are discussed. The Bsc thesis can be divided into two main sections: one of them is the research of classical statistical methods, while the other is dealing with artificial neural networks.
I start with the introduction of statistical methods based on literature survey. Firstly structures of regression models and some well-tried regression methods used and working in practice are shown. After giving a brief survey of the biological neural networks artificial neural networks are introduced. The most important properties of the structure of neural networks are presented. Then, the structure of multi-layer perceptron (MLP) mostly used in short-term load forecasting is shown. Finally, the training of MLP, and the most important questions of construction are discussed. Networks with radial basis functions are mentioned but not discussed in details because during the realisation I have created and tested an MLP-type network. Moreover, some artificial neural networks and learning algorithms tested in practice are introduced.
During the implementation I aimed to create forecasting methods with the least possible errors. To evaluate the quality of the models MAPE was used, like in many other publications and studies. At first, three different regression models were constructed by EXCEL: forecasts of „Similar Day”, „Weather Adaptive” and „Pattern Matching”, then these models were compared. After that, I implemented a feed forward MLP-type neural model and tested it in MATLAB. Finally, I evaluated the test results of the models constructed.