A Data-based Mechanism Learning Paradigm for Algebraic Relationships in Power Systems

The increasing proportion of renewables poses significant challenges to the safe and stable operation of power systems. Inspired by the concept of foundation models, represented by GPT, the construction of foundation models suitable for power systems offers new perspectives for solving these challenges. This paper proposes a novel learning paradigm for algebraic relationships, which has the potential to construct foundation models for power systems. With masked modeling techniques, the paradigm randomly masks a portion of the power flow results and forces the model to recover the missing parts based on the known information. This process departs from the task-driven perspective, enabling the model to learn the complete algebraic relationships. We also design a Graph Transformer-based architecture to be compatible with this paradigm. Experimental results demonstrate that the proposed method not only satisfies the algebraic relationships but also exhibits strong generalization capabilities.