Abstract: The coordinated optimization of “source-networkload-storage”in distribution network is an important means to absorb renewable energy,in which reactive power optimization can ensure the safe,reliable and economic operation of the system. This paper proposes a reactive power coordination method for “sourcenetwork-load-storage” system of distribution network based on deep learning and intelligent online scene matching. It fully considers the characteristics of operation scenes,directly uses big data of distribution network operation to generate the offline historical scene library,and constructs the offline historical strategy library based on multi-objective particle swarm optimization algorithm. Then,the historical scene library and real-time scenes to be optimized are classified twice based on different optimization objectives and K-means clustering algorithm. Next,the historical scene library is taken as the training set,and the real-time scenes to be optimized are taken as the test set. Intelligent online scene matching is realized based on the deep neural network technology.The paper determines the success or failure of the matching according to the matching evaluation index and allocates the reactive power optimization scheme by matching the historical strategy or online optimization. Finally,random loads such as distributed photovoltaics,energy storage devices and electric vehicle charging stations are added into IEEE 30-bus simulation model to verify the calculation example. The results show that the method is flexible and effective for coordinated reactive power optimization,independent of models and parameters,and greatly improves the decision-making efficiency.