The electromagnetic transient simulation of large-scale power system is often accelerated by parallel simulation with sub-networks. However, the general network partitioning method based on geographical location information is not self-adaptive, resulting in seriously constrained sub-network numbers, uneven scales, and many associated lines. This paper proposes a self-adaptive network partitioning algorithm based on the k-way thought, making use of the high sparsity of power system. This algorithm combines the mapping relationship between the original data and the graph structure of power system, the preprocessing skills of simplifying complex graph structure and the linear deterministic greedy algorithm to overcome the shortcomings of general network partitioning method and further improve the speed of electromagnetic transient parallel simulation of power system. The algorithm’s results are applied to the electromagnetic transient parallel simulation, verifying its reliability and acceleration effect.