As new energy sources are connected to the power system and sent out through DC, the transient overvoltage problem of the sending-end system is gradually becoming prominent. Therefore, a graph convolutional network (GCN)-based transient overvoltage evaluation model is proposed to quickly and accurately estimate the transient overvoltage severity at each DC near-zone node in expected disturbance scenarios such as DC block and commutation failure. This model takes the state parameters and the network topology before a DC fault occurs in the grid as input features, and can predict the transient overvoltage severity of multiple critical nodes of the grid (e.g., wind farm aggregation node) simultaneously. A case study using a two-region system with cross-region DC asynchronous interconnection verifies that the model can be adapted to different grid operational modes, such as multiple grid topologies and different new energy generation ratios, and has a strong generalisability. At the same time, the proposed model reveals the key factors that have the greatest impact on overvoltage severity, and has a certain interpretability, which can provide effective guidance for the prevention and control of transient overvoltage.