Abstract: The rural distribution network is usually connected to the distributed power supply with large capacity, such as small hydropower, which causes a lot of line loss in the process of reverse power transmission to the superior grid. Therefore, this paper proposes a real-time scheduling strategy to reduce the line loss of the distribution network. Firstly, considering the active and reactive power control of distributed generation and the control of on-load tap changer, a loss reduction model of distribution network scheduling is established based on the branch power flow model; secondly, by constructing a high-dimensional random matrix, the characteristics that can represent the operation state are extracted from the operation time series data of distribution network as the input, and the historical regulation strategy of distribution network is thermally coded as the output. Then, the deep bidirectional long and short-term memory network（BI-LSTM） is used to learn the function mapping between the characteristics of distribution network and the network loss reduction strategy, a real-time loss reduction model of active distribution network based on data deep learning driving is established. Finally, based on the actual active distribution system simulation, the simulation results show that the proposed real-time scheduling algorithm can optimize the output curve of small hydropower, improve the local absorption rate of distributed power, and reduce the network loss on the premise of ensuring the income of small hydropower network.