Abstract: The existing EV charging load forecasting studies are mostly conducted on a single forecasting object. Meanwhile, less research has been conducted on the spatial-temporal forecasting of charging loads at multiple public charging stations. The charging loads of public charging stations fluctuate drastically and are more difficult to forecast compared to those at private charging facilities. To this end, an adaptive spatial-temporal graph neural convolutional network based spatial-temporal short-term forecasting method for charging loads at multi-public charging stations is proposed. First, multi-node feature sets are constructed by using Rapid-MIC. Through data adaptive graph generation, a similar-weighted spatial-temporal graph is constructed to reconstruct the spatial connection relationship of multi-public charging stations. Then, graph convolution layers are constructed to generate the spatial aggregation features based on the differentiated coupled spatial-temporal correlations of each node, so as to realize the differential feature enhancement of nodes in the whole domain. Meanwhile, the charging patterns of different nodes are learned by node adaptive parameter learning. Finally, the temporal domain features of spatial aggregation features are mined by gated recurrent unit layers. The symmetric mean absolute percentage error(SMAPE) and mean absolute error(MAE) values of the proposed spatial-temporal forecasting method for charging loads at public charging stations are 12.95% and 31.72 kW.