In response to the challenges such as spatial heterogeneity, complex spatio-temporal dependence and multi-scale temporal characteristics existing in urban carbon emission prediction, this study proposes a spatio-temporal adaptive graph convolutional recurrent network (ST-AGCRN). The model breaks through the limitations of the traditional prediction paradigm through a triple innovation mechanism: Firstly, a dynamic graph generation module is designed, and an adaptive adjacency matrix is constructed using a learnable feature mapping function to explicitly quantify implicit spatial correlations such as industrial chain collaboration, solving the problem that the static graph structure cannot represent the dynamic changes of economic correlations. Secondly, innovate a multi-stage spatial feature fusion architecture. Through dual-channel diffusion convolution, collaboratively integrate the carbon transmission paths dominated by geographical proximity and economic correlation. Combine the gating mechanism to balance the contributions of dynamic and static graphs, and achieve collaborative modeling of local diffusion and global transmission. Finally, a hierarchical time series analysis framework is constructed, integrating the dilated perception capability of gated temporal convolution with the dynamic coupling mechanism of spatio-temporal perception cyclic units to effectively capture the interaction effects of policy intervention, seasonal cycles, and meteorological mutations. Experiments in the three major urban agglomerations show that the model significantly improves the prediction accuracy of industrial cities, enhances the robustness of extreme meteorological scenarios, and optimizes the training efficiency.