There are two main sources that produce anomalous load profiles and introduce error in Short Term Load Forecasting in power systems, anomalous situation in power system such as various contingencies of overload and interruptions, imposed load shedding, voltage collapse etc. and the shortcomings of operators in manual recording data without error. In this thesis with precise examination of load profiles in a real power system, various methods of filtering anomalous data is used and a new method based on Principal Component Analysis (PCA) is suggested that can be used efficiently to recognize unusual profiles. Then, a Short Term Load Forecasting system (STLF) is designed according to the Feed Forward Neural Networks (FFNN) and its various parameters are optimally tuned by error sensitivity analysis. To fulfill minimum error and simplicity of implementation, several models are presented and the model with proper precision and less error is selected. It is shown that elimination of temperature input does not make considerable error in the forecasted loads. Independence of the algorithm to temperature data can be considered as an advantage of the procedure. Special load profile and small number of patterns in data cause high error rate in forecasting the load profile in holidays. Accordingly in this thesis, to forecast load profile for these days, a proper procedure is presented. The result indicates that it has acceptable performance and reasonable precision.
