A prediction method for mold breakout based on feature vectors and hierarchical clustering is disclosed, which comprises: respectively extracting temperature feature vectors of historical data under sticking breakout and normal conditions and on-line actually measured data to establish a feature vector sample set; performing normalization and hierarchical clustering on the sample set; and checking and judging whether the feature vectors extracted on line belong to a breakout cluster, and then identifying and predicting mold breakout. The method avoids the steps of tedious adjustment and modification of alarm threshold and other parameters, overcomes the artificial dependence of the previous breakout prediction method, has good robustness and mobility; and through temperature feature extraction, achieves accurate identification of sticking breakout temperature patterns, avoids missing alarms and significantly reduces the number of times of false alarms, and greatly reduces the data calculation amount and calculation time, guaranteeing the timeliness of on-line prediction.