Load current data can be regarded as stochastic disturbance of voltage. Three decompositions and re-construction of wavelet package method are used to analyze load modeling data， and the characteristic vectors of load data are accurately constructed which are used to classify load data. Load data are classified by fuzzy subtraction clustering based on the standardized characteristic vectors. The validity and veracity of the method proposed have been proved by characteristics extraction and clustering for dynamic lab data and transformer substation data. The high precision and convergence make it significant for characteristic extraction and clustering of bulk modeling data.