The total number of contingencies needed by early-warning of transient stability is huge. In order to reduce the computation burden in every analysis cycle， an adaptive contingency screening method， which is designed to filter out slight cases as much as possible and keep all the severe ones for further detailed assessment， is presented. Based on the stability margins and dynamics modes of contingencies resulting from the last assessment cycle， the critical cases to system stability are firstly identified by using quantitative analysis method. Secondly， based on the changes in power flow on the critical elements， stability margin thresholds for contingency screening are identified. Finally， those contingencies need further detailed analysis for new cycles can be determined by comparing the prediction value and the relevant thresholds. Real engineering applications in actual power grid have fully confirmed the validity and practicability of this method.