Multiplex Graph Transformer Boosts Power Grid Model Generalization.

Graph neural networks (GNNs) often struggle with power grid problems when the grid's topology changes, a phenomenon termed "topology overfitting." This occurs because models trained on specific topologies encode relational structures unique to those grids, failing on unseen configurations. To address this, a new model called MxGPS (Multiplex GPS) has been developed. MxGPS is a multiplex graph transformer that uses multiple task-specialized branches running over a single shared node encoder. It's jointly trained on Static State Estimation (SSE) and AC Power Flow (PF) tasks, using a self-supervised pre-training and multi-task fine-tuning approach. This joint objective forces the shared encoder to learn underlying physics rather than topology-specific patterns. The model achieved a 0% boundary violation rate on four unseen power grid topologies, significantly outperforming other models that degraded severely under topology shifts, all while using 12 times fewer parameters than a reference baseline.