With the increasingly prominent role of virtual power plant technology under the background of “double high”

new advanced scheduling applications increasingly rely on the accuracy and integrity of basic data. In this paper

we propose a privacy-preserving reconstruction method for virtual power plant load data based on federated learning and multimodal bidirectional encoder representations from transformers (BERT). Firstly

an edge-centralized privacy-protected training model is proposed based on federated learning. This model reconstructs missing data by leveraging contextual information accurately through a bidirectional encoding mechanism. Secondly

a highly coupled associative multi-modal feature input method is proposed

which considers the timing and physical characteristics of the electricity price and load power data

and constructs a BERT bidirectional model to improve the dimensionality and extraction speed of the input data

and improve the reconstruction accuracy and efficiency. Finally

more perfect information features of load active and reactive power data are captured

and the optimal value points are selected based on the “cross point” identification

and the cross-modal learning ability is used to improve the reconstruction accuracy. Simulation results show that the FLN-BERT-PIN method improves both the accuracy and robustness of load data reconstruction

while providing a practical pathway for advancing virtual power plant (VPP) data governance

intelligent dispatch

market operations

and privacy protection in future power systems.