Abstract: With the transition to the new power system, renewable energy sources, particularly photovoltaic (PV) generation, have become the dominant power supplier. Accurate forecasting of PV output is critical for maintaining power system stability. However, most existing forecasting methods have mainly focused on temporal sequence modeling while overlooking the spatial feature information among PV clusters. To address this issue, this study proposes a distributed PV short-term forecasting model that integrates fuzzy C-means (FCM) clustering and spatiotemporal graph neural network. Firstly, distributed PV stations are clustered with the FCM algorithm, and then a graph structure is constructed. The graph attention network (GAT) is employed to capture spatial features among distributed PV stations, and the sequences are processed by the bidirectional gated recurrent unit (BiGRU) to extract temporal features. In addition, the grey wolf optimizer (GWO) is introduced to optimize the model hyperparameters. The experimental results based on real data from 10 distributed PV stations verify that the proposed forecasting method effectively integrates spatial and temporal features of distributed PV stations and improves the prediction accuracy by at least 6.7% compared with other models.