Applications Of Soft Computing And Data Mining In Electrical Load Forecasting

Data mining distills connotative knowledge and information from abundant data, while soft computing is an effective mothed to establish intelligent computing systems. This dissertation combines the two motheds to accomplish two main tasks: outliers processing in load data and multi-factor STLF (short term load forecasting) system modeling. The algorithms operate based on data completely, and are more intelligent and scientific.The dissertation layes stress on two main thoughts of data mining: classification and cluster analysis, some methods of which are discussed.Neural networks and fuzzy systems lay the two main foundations of soft computing. The principles, constructures and algorithms of BP and Kohonen network and the TSK fuzzy model are discussed. The two neural networks are combined to processing the outliers in load data, while the fuzzy model is the kernel of a multi-factor load forecasting system.The outlier identification is divided into two sequential parts: the robust day-load-curves cluster and the bad curve pattern classifcation. By analysising the effects of Kohonen network clustering and BP network classifcation, the dissertation designs an outlier identification model comprising these two kinds of neural network and implements the tasks of bad data identifications and adjustments.A STLF system is modeled based on ANFIS (Adaptive Neural-Fuzzy Inference System). Two main problems are sovled: fuzzy system structure identifications and ANFIS parameters identifications. A decision tree classifier is used to deal with the first task, which can find the load pattern preliminarily, and reduces the number of parameters to be adjusted. The Quasi-Newton method is applied to solving the second problem, which is suitable to optimizing large number of parameters.Difficulties of fuzzy modeling are the variables selection and the input space division in the fuzzy structure identification. CART can eliminate the unnecessary variables, and then divides the input space as a tree-shape. Appropriate membership functions improve the convergency of the parameters identification.