The invention discloses a three-stage multi-visual-angle characteristics integrated electric power theft classification predicting method. The method comprises: first, using the to-be-analyzed electric power consuming data of the clients as a training set; filling the missing data of the current day's consumed electric power amount, the current day's meter read number, and the last day's meter read number with -1 and 0; and forming two kinds of pre-processed data; second, extracting the characteristics of the each kind of pre-processed data from different visual angles; combining the extracted characteristics of all the visual angles; using a plurality of different machine learning algorithms for classification predicting to process so as to obtain the electric power theft probability among the clients in the training set and the testing set; and finally, respectively using a linear model and a tree model to predict the output of the second stage; then, seeking the mean value and obtaining the to-be-predicted electric power theft probability. Based on the currently available pile model integrated learning method, the diversity of data, the diversity of models and over-fitting processing are added so as to realize more accurate predicting on the electric power theft probability among the clients.