پالايش داده هاي آموزشي شبكه عصبي و بررسي تأثير آن در كاهش خطاي پيش بيني كوتاه مدت بار سيستم هاي قدرت

Filtering out the Training Data of ANN and Evaluation of Its Effect on Error Reduction in Short-Term Load Forecasting in Power Systems

Occurrence of special events such as overloads, load shedding, interruptions and faults, and shortcomings of operators in an error free manual recording of data are the main sources of anomalous load profiles imposing errors in Artificial Neural Network based Short-Term Load Forecasting (ANNSTLF) systems. In this paper, an ANNSTLF system is properly designed for the real power system of Isfahan Province and used to compare the impacts of filtering out the anomalous data by expert experience, statistical analysis, and a new algorithm based on the Principal Component Analysis (PCA). The results show that filtering out anomalous load-profiles before ANNSTLF can reduce the forecasting error efficiently. It is also shown that the proposed PCA filtering method is simpler in application and faster in response, yielding accurate forecasting results