The spatial heterogeneity of photovoltaic(PV) clusters makes it difficult to unify the weather types among different sites. Existing daily-scale weather classification methods ignore intraday transitional weather characteristics

making it difficult to provide effective scenario information for forecasting model construction. To address these issues

this paper proposes a short-term power forecasting method for PV clusters considering weather inconsistency and intraday variations in power generation characteristics. First

representative stations from each sub-cluster within the PV cluster are selected

and their similar-day power profiles are calculated. Second

based on the similar-day power profiles of the representative stations

a clustering model is employed to divide the output levels at different time points

where continuous output levels represent different power generation scenarios. Finally

based on the extraction of dynamic spatial correlations using a graph convolutional network(GCN)

output-level information is utilized to guide the variation of edge weights between nodes. Meanwhile

an interpretable time-fusion Transformer is adopted to construct a unified power forecasting model under different output scenarios. Through structures such as the variable selection network and static variable encoding

the output-level context vector is integrated throughout the modeling process. By dynamically selecting features under different scenarios

the proposed method further achieves scenario information-guided forecasting model construction and interaction with scenario characteristics. The interpretability of the model is enhanced by visualizing variable weights. The effectiveness of the proposed method is validated using data from a PV cluster in Inner Mongolia Autonomous Region.