Abstract: The wide access of the high proportion of power electronic equipment and the high proportion of distributed photovoltaic power, and the improvement of the urban cabling rate make the reactive power characteristics on the user side of the distribution network complicate. The increased uncertainty of the load reactive power consumption is not conducive to the safe operation of distribution network. Therefore, to better optimize reactive power, the combined scenarios and their probabilities of daily power-factor variation curves of different loads are used to reflect the uncertainty of reactive power. Taking the minimum expected value of operation cost as the objective function, an optimal configuration model for the expected value of multiple reactive power scenarios is established. Firstly, multiple one-dimensional convolutional autoencoders（1 D-CAEs） is used to extract the low-dimensional representation of the daily power factor data of different users. Then, the k-means method is used for scene reduction to obtain typical daily power-factor variation scenes, and multi-user scenario set is combined. Finally, the expected value reactive power optimization model is established, and the particle swarm algorithm is used to solve it to determine the optimal configuration scheme. According to the reactive power consumption scenarios of users in a distribution network in Shanghai, the modified IEEE 33-node system is taken as an example to verify the effectiveness of the proposed method.