Abstract：As the proportion of installed photovoltaic（PV）capacity increases year by year，accurate prediction

of PV output and the realization of group control and management of PV clusters become crucial. A multi-site PV

output prediction method that integrates deep fusion of graph convolutional neural network（GCN），long shortterm memory（LSTM），and temporal pattern attention（TPA）was proposed. First，the input features of multi-site PV output time series curves and numerical weather forecast data were transformed into a graph structure to establish a GCN-LSTM model，which extracts the hidden spatio-temporal dependencies among PV clusters.

Second，an attention mechanism was introduced to weight and correct input data features，enhancing the value of

key data. Then，based on the spatio-temporal prediction results of the PV clusters，dominant nodes that sensitively

reflect the voltage changes of the cluster were selected，and an inter-cluster coordinated optimization strategy was

constructed with the goals of ensuring no voltage limit violations in the entire regional nodes and minimizing the

system's network losses. Following that，an intra-cluster autonomous optimization and control strategy was

constructed within the cluster based on the coordinated optimization strategy results，aiming for the safe operation

of cluster node voltages，minimum cluster network losses，and maximum local consumption of distributed PVs.

Finally，simulation results of actual multi-site PV cluster output data show that the proposed method can efficiently

extract the spatio-temporal correlations between different PV stations，reduce the prediction error of PV output，and effectively improve the safety and economy of PV clusters.

      Key words：PV output prediction；graph convolutional neural network（GCN）；adaptive adjacency matrix；

temporal pattern attention（TPA）