An approach of angle stability assessment is proposed considering the characteristics in fast refresh rate and large amount of WAMS data. First, initial set of features is selected and then reduced by kernel principal component analysis to filter redundant features. Then, training sample aggregation is constructed and initial set of features of each sample are calculated and reduced. Finally, the method is evaluated through training ECVM classifiers, the accuracy of which is validated by test data set. Simulations on New England 10-machine 39-node system show the proposed algorithm has higher classified accuracy, and compared to traditional methods it has lower evaluation time for a single sample and engineering application value.