The Application Research On Short-Term Load Forecasting Method Based On Knowledge Base

The key to improve the whole load prediction accuracy is properly handling the prediction problem in abnormal days. A novel approach for Short-Term Load Forecasting (STLF) based on knowledge base technology is put forward in the dissertation. First of all, through adjusting amplitude of wavelet modulus maxima and processing the wavelet decomposed detail signal by soft threshold based on wavelet analysis and singularity theory, fault data in original loads are eliminated. Then, the knowledge base filled with cases is constituted with processed load data and influential factors which are organized in according to case presentation. The k-nearest-neighbor is applied to find the most similar cases whose cases attributes are nearest with the cases to be predicted, and the attributes weights of influential factors are calculated by weights computing based on rough set. In the process of extraction, information entropy and principal component analysis are integrated to reduce similar cases set. The processed load sequence from similar cases is used to train the BP neural network based on Levenberg-Marquardt dynamic numerical optimization algorithm. Aiming at abnormal days, data mining technology based on decision-tree is used to construct revision model so that to make a further revision on forecasting. Finally, the forecasting results are preserved in cases set in according to case presentation as refreshment resources. The testing results of STLF in actual power network show that the proposed method aiming at abnormal day possesses higher forecasting accuracy and better adaptability.