Abstract: In the face of the increasingly complex challenges posed by loads to the operational safety of distribution networks, the existing early warning methods based on a single temporal or spatial perspective are difficult to comprehensively reflect the spatio-temporal dynamic characteristics of power outage risks as the distribution network topology expands. To this end, a dynamic early warning method for power outage risk in distribution networks based on spatio-temporal graph convolutional network (ST-GCN) is proposed. We model the distribution network as a graph, define its spatial structure with nodes such as busbars and load points, and edges such as lines and switches, and enrich the spatio-temporal dynamic characteristics of its nodes with multi-source data such as SCADA data, meteorological data and historical loads, while simultaneously mining the complex dependencies in space and time. The case results show that, compared with the three benchmark methods, the proposed method has obvious advantages.