Abstract: There are new challenges for the traditional reactive voltage control in computing speed and accuracy with the high penetration of distributed generation, energy storage and flexible load in the distribution network. In this paper, a data-driven reactive voltage control method is proposed for the control of the reactive voltage by tracking the operating parameters of the actual system. An auto encoder is combined into the extreme learning machine to construct a deep learning mechanism, and the direct coupling relationship between the input and output of the extreme learning machine is established based on the automatic encoder. The unsupervised learning and the supervised learning are comprehensively integrated to reduce the iterative process of the training model in the deep extreme learning machine. Then, a series of distribution network operation scenarios based on the distributed power supply and flexible load prediction information are built using the Monte Carlo method. The deep extreme learning machine is used to mine the internal relationship between the optimal operation in the operation scenarios and the status of the reactive voltage regulators. The mapping relationship between the grid operation scenarios and the reactive power voltage regulation strategies of the system is established. The proposed method realizes the tracking of the distribution network operation state and the optimal scheduling of reactive power regulation equipment without depending on the actual system power flow calculation, providing a support for the reactive voltage control.