Abstract: Large-scale doubly-fed wind farms which are connected to the power grid through series capacitor compensation can easily lead to subsynchronous oscillation of the power system. If the oscillation phenomenon can not be timely detected to give an accurate alarm, it will seriously threaten the safe and stable operation of the system. The subsynchronous oscillation early warning is based on the online measurement data of the power system to judge the stability of the system oscillation and provide real-time and reliable early warning information for dispatchers. In order to solve the problem that the existing online monitoring of subsynchronous oscillation cannot realize pre-warning, this paper proposes a subsynchronous oscillation early warning method based on graph attention network (GAT). First, the key influencing factors of secondary synchronous oscillation are screened from the wind farm side and the grid side, respectively, to reduce the redundancy of input feature information. Secondly, based on the multi-head attention mechanism, a multi-head attention network is constructed to realize the aggregation of sub-synchronous oscillation characteristics among different wind farms after considering the differences in the coupling effects between different wind farms and between wind farms and the power grid. Finally, the subsynchronous oscillation stability early warning is carried out on the multi-wind farm integrated grid system, and the accuracy and anti-interference of the method proposed in this paper are verified.