With the large-scale integration of distributed photovoltaic (PV) generation, power quality issues in distribution networks have become increasingly pronounced. In particular, overvoltage problems can cause severe damage to electrical equipment and even lead to system-wide outages. Therefore, the ability to anticipate voltage conditions at individual PV user sites is of critical importance, as it enables grid operators to formulate appropriate control strategies in advance and effectively reduce operational risks. Traditional forecasting methods face limitations in simultaneous modeling spatiotemporal interdependency and numeric weather predictions (NWP) correlations. To address these challenges, this study proposes a weather-aware spatiotemporal graph neural network for multi-step three-phase voltage prediction across multiple PV users, termed AST-CA-GNN. Building upon the ASTGNN framework, the model innovatively introduces cross-attention to enhance sensitivity to weather information, and constructs a dynamic graph convolution module based on prior spatial topology graph to capture time-varying spatial correlations between PV users. The experimental data are sourced from a 380V rural distribution system in China, including historical PV power observations, distribution network topology information, and NWP variables. Experimental results show that the proposed model outperforms other baseline models across all evaluation metrics in the three-phase voltage forecasting task, and achieves the best performance in the voltage limit violation prediction.
