The integration of high-penetration renewable energy sources into power grids poses significant challenges for photovoltaic (PV) cluster power forecasting

including quantitative modeling of complex spatial correlations

spatial propagation of power fluctuations

and characterization of technical heterogeneity

this paper proposes a short-term collaborative forecasting method based on a graph state space model (GSSM). The proposed method commences with multi-scale graph construction

which adaptively incorporates critical information

including geographical proximity

technological homogeneity

and electrical coupling among PV power plants to precisely capture dynamic inter-plant correlations. Subsequently

a continuous-time state-space formulation is developed

incorporating an explicit graph diffusion term to physically interpret the propagation of meteorological disturbances across the cluster. To enhance responsiveness to transient weather events

a selective scanning mechanism is integrated

enabling dynamic modulation of model parameters

and further achieving short-term collaborative prediction for photovoltaic clusters through a multi-step predictor. Empirical evaluations on the publicly available PVOD dataset demonstrate that GSSM substantially surpasses baseline models

including long short-term memory (LSTM) networks and spatio-temporal graph convolutional networks (STGCN)

in training efficiency

predictive accuracy

and long-term stability. Moreover

GSSM exhibits robust generalization to heterogeneous PV cluster configurations. This research can provide technical support for power grid dispatch and assist in the safe operation of high-proportion renewable energy systems.