With the large increase of distributed generation and intermittent loads, the conventional reactive power optimization intelligent algorithm needs to further improve the convergence efficiency and calculation speed while ensuring the optimization accuracy. As a result, an improved reactive power optimization method using similar search and threshold judgment based on large sample data is proposed. The method uses the similarity search algorithm to search the database for the historical load that is most similar to the current load, and uses the reactive power optimization plan of the historical load as the target scheme. Through the similarity threshold judgment, after the local correction by the expert knowledge method, the target scheme with higher similarity is directly called. When the similarity is low, the target scheme is added to the particle swarm optimization algorithm as the “optimal particle” to perform fast optimization and correction, which improves the speed and accuracy of the conventional intelligent algorithm. Finally, the regional power grid and IEEE8500 example are used to verify that the method has excellent optimization effect and optimization speed.