Photovoltaic (PV) output forecasting is crucial for optimizing power grid dispatching and enhancing new energy consumption, especially with the rapid development of the PV industry in China. To capture the spatial correlation among different arrays in a PV site and the temporal dynamics of PV power outputs, a novel ultra-short-term PV output forecasting method based on a graph convolutional network and long short-term memory (GCN-LSTM) network is proposed. The proposed method first constructs a graph model to represent the connection relationships of different arrays on the PV site. Then the graph convolutional network is used to extract spatial features from the graph model to obtain time series data that incorporate the spatial relationships among different arrays. Finally, time series data is input into the LSTM network to perform PV output prediction. Experiments demonstrate that the GCN-LSTM-based PV output forecasting method achieves high accuracy and stability, which makes up for the inherent limitations of prediction methods based on time series data and shows promising application potential in large-scale power plants.