The present invention provides a transient assessment evaluation feature selection method and device. The method comprises the following steps that: a weighted random forest model is constructed; features with the lowest importance are deleted from calculated feature importance; final feature sequencing is performed on the retained technical features; and features which are high in ranking are retained according to OOB (Out-Of-Bag) error rate curve changes. According to the transient assessment evaluation feature selection method and device provided by the technical schemes of the invention, few unstable sample data weight factors are introduced into the weighted random forest, so that the influence of unbalanced data on feature selection can be decreased, feature subsets of which the performance can be better than that of feature subsets selected by an ordinary random forest algorithm can be selected; and threshold values are not required to be set manually, and the rationality of the selected feature subsets can be ensured.