Cluster Forecasting of Multi‐Site Photovoltaic Power Generation Using Virtual Collection and Transfer Learning

ABSTRACT Accurate photovoltaic (PV) power cluster forecasting is critical for optimizing spinning reserves and guaranteeing the security of power supply. However, practical implementation faces three interconnected bottlenecks: weak spatial‐temporal modeling, vulnerability to data interruptions and high computational overhead when scaling to new sites. To address these challenges systematically, this paper proposes a unified forecasting framework integrating multi‐channel deep learning, virtual data collection and transfer learning. First, a “multi‐channel + fully connected” CNN‐LSTM architecture is utilized to extract site‐specific features and dynamically synthesize cluster‐level predictions. Second, for missing data scenarios, a LightGBM‐based virtual collection module efficiently imputes gaps using real‐time data from spatially correlated neighbouring sites. Finally, a parameter‐freezing transfer learning strategy enables rapid model adaptation for newly integrated sites. Validated on 21 PV sites totalling 1075 MW in Gansu Province, the framework demonstrates significant improvements. The proposed cluster approach reduces forecasting errors by 0.5% to 0.8% compared to traditional single‐site baselines. Furthermore, the transfer learning module accelerates model training by approximately 82% while maintaining comparable forecasting accuracy. This highly scalable framework provides a practical, data‐driven tool for system operators to mitigate grid integration risks.