Abstract The massive perception data based on efficient analysis and intelligent decision have put forward higher requirements for high‐precision time synchronisation with the construction and development of smart power grid. However, multi‐reference source time‐frequency synchronisation of power system only selects the best method after comparison, which cannot make the most efficient use of the existing resources. It also cannot meet the need for high‐precision time synchronisation of future power system. The existing multi‐reference source synthesis algorithms cannot take into account both long‐term stability and high‐precision synchronous output. This article presents a multi‐reference source weighted improved noise model and the high‐precision output method. The multi‐reference source error after classification is eliminated by leading into classification vector and classification coefficient. The synthesised frequency offset or the time precision of output can be optimised as the objective function by weighted classification algorithm and genetic algorithm. A simulation example based on the synthesis of two satellite system clock sources and three local caesium reference sources shows that the peak value of long‐term output accuracy is controlled within 10 ns after classification weighted synthesis and optimisation, which is better than that of any single reference source.