The increasing penetration of renewable energy and power electronic devices in modern power systems has led to prominent low-inertia characteristics. The frequency instability risk has increased dramatically after a large-capacity active power disturbance event. For this reason

it is necessary to carry out an online

real-time system frequency security assessment with respect to the anticipated disturbances during the steady-state operation of the power system to discover the potential frequency security risk in time and guide the subsequent frequency security preventive control. This paper proposes a Transformer-Graph convolutional network (Tr-GCN)-based online intelligent model for frequency security assessment and designs a hybrid input feature set containing system steady-state operation state information and anticipated disturbance information. It can be applied to frequency response index prediction under various steady-state operation conditions and anticipated disturbances. The model encodes and calculates the disturbance location and amounts more rationally based on the power system topology. In addition

by introducing a multi-head self-attention mechanism

the model selectively focuses on more meaningful information to improve the model expression capability and prediction accuracy. The IEEE-39 bus system with renewable energy and the Northeast China power system (2

131 nodes) are examples for the case studies

which verifies the proposed method's accuracy

effectiveness

and scalability.