Abstract: Rolling early warning can identify future insecure operation status of power grids and provide decision-making guiding information, which is an effective technique to resist systemic security risks. Considering the uncertainties on source and load sides,a data-driven graded rolling early warning method of dynamic security for large-scale hybrid AC/DC power grid is proposed. First,a control cost based early warning grading strategy is proposed and the control cost of typical control measures is analyzed. The early warning levels are distinguished according to the type of control measures that are needed to ensure the system security.Then, a coordinated multi-time-scale rolling early warning framework is established. Using the characteristic that the prediction accuracy on source and load sides increases as the time to be predicted approaches, the accuracy of the early warning results is improved and it is matched with the control measures of different action sequences. Finally, an incremental learning based online update method of machine learning model is proposed, which uses dynamically added sample data to improve the accuracy of model evaluation, and trains multiple models through replacement sampling to calculate the credibility of early warning results within a given error tolerance. Simulation results of a real power grid demonstrate that the proposed method can effectively track the dynamic security risk situation, and guide the matched control decisions.