ABSTRACT
This chapter explains all phases involved in the development of an artificial neural networks-based shortterm load forecasting (STLF) system. It describes the first stage of an input variable selection process. Input space representation is probably the most important subtask in load forecasting. STLF can be enhanced by data preprocessing. The basic motivation for normalizing input and output variables is to make them comparable for the training process. The objective of wavelet filtering is to identify different sources of useful information embedded in a load time series. Mutual information has been used for selecting inputs to neural networks. The underlying motivation for its application is the capacity for detecting high-order statistical relationships among variables. Autonomous means that both input selection and model structure identification are performed in an adaptive and automatic mode. The input representation problem when applying neural network models to forecast electric loads has been overlooked for a long time.
