Abstract: The dynamic frequency prediction method based on machine learning generally ignores the changes of system topology, causing the result that the original trained model may not be applicable. To solve this problem, a new method for frequency prediction based on random forest(RF) is proposed. Considering that the training time of the RF algorithm is proportional to the number of the features, a Spearman correlation and hierarchical clustering aggregation method is used to remove the redundancy and reduce the number of the initial features. On this basis, the features that play a key role in the system frequency are extracted, which further reduces the training time. The different situations before and after the redundancy removal and the key features extraction are compared and analyzed, showing that the proposed algorithm may greatly shorten the training time under the premise of ensuring high prediction accuracy. Simulation results on the New England 39 system show that the proposed algorithm is fast, fault tolerant and accurate.