PhysBridge: Response-supervised grid pretraining with physics-constrained local correction for transferable power-system intelligence
- 1State Grid Shandong Electric Power Research Institute, Jinan, Shandong, China.
- 2Shandong Key Laboratory of Energy Industry Internet Big Data Technology, Jinan, Shandong, China.
- 3State Grid Shandong Electric Power Company Linyi Power Supply Company, Linyi, Shandong, China.
Abstract
Operating modern power grids requires coordinated solutions for optimal dispatch, state estimation, price prediction, and contingency analysis, yet these tasks are usually modeled separately. We present PhysBridge, a response-oriented pretraining framework in which historical windows provide context while solver-generated perturbation responses, active constraints, and local sensitivities define the pretraining targets. PhysBridge combines a temporal-graph encoder with a Primal-Dual Response Bridge (PDR-Bridge) that performs safeguarded local refinement and outputs transferable primal, projection-dual, sensitivity, and health tokens. Across IEEE 14-, 30-, 57-, 118-, and 300-bus systems, the benchmark contains 26.3 million solver-generated perturbation-response pairs; in the 300-bus transfer setting, pretraining uses only the 21.04 million nominal pairs from the 14-, 30-, 57-, and 118-bus source topologies. Across three runs on unseen IEEE 300-bus transfer with 10% labels, PhysBridge obtains a 2.21±0.05% OPF cost gap, 96.3±0.2% feasibility, 1.84±0.05 $/MWh LMP MAE, and 91.6±0.3% contingency/fault event macro F1. Relative to a matched rollout-only model trained on the same source corpus without response targets or bridge refinement, these values correspond to a 50.8% lower OPF gap and a 7.4 percentage-point higher event F1. The bridge requires 12.4 ms per 300-bus forward pass and activates fallback on 4.6% of stressed windows. Comparisons with supervised-only OPF specialists that do not receive matched solver-generated pretraining are interpreted as task-relevant performance under the proposed pretraining regime, not as architecture-only superiority.
Related
- Biobank-scale self-supervised genotype pretraining yields transferable representations for human genetics
- Self-Supervised Pretraining of Molecular Graph Encoders with LeJEPA
- MtReNIRS: A Self-Supervised Foundation Model for Transferable Neurovascular Representations
- URCHIN: A Horizontal Spiking Language Model for Data-Constrained Pretraining
- BiomECG: Learning Transferable ECG Representations from Small Cohorts
BibTeX
@article{liu2026physbridge,
title = {{PhysBridge: Response-supervised grid pretraining with physics-constrained local correction for transferable power-system intelligence}},
author = {Donglan Liu and Shanjie Xu and Yuhui Jin and Honglei Yao and Qinguo Liu},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0359953},
url = {https://doi.org/10.1371/journal.pone.0359953},
note = {CC BY},
}
RIS
TY - JOUR
TI - PhysBridge: Response-supervised grid pretraining with physics-constrained local correction for transferable power-system intelligence
AU - Donglan Liu
AU - Shanjie Xu
AU - Yuhui Jin
AU - Honglei Yao
AU - Qinguo Liu
PY - 2026
DA - 2026/10/28
JO - PLOS One
DO - 10.1371/journal.pone.0359953
UR - https://doi.org/10.1371/journal.pone.0359953
AB - Operating modern power grids requires coordinated solutions for optimal dispatch, state estimation, price prediction, and contingency analysis, yet these tasks are usually modeled separately. We present PhysBridge, a response-oriented pretraining framework in which historical windows provide context while solver-generated perturbation responses, active constraints, and local sensitivities define the pretraining targets. PhysBridge combines a temporal-graph encoder with a Primal-Dual Response Bridge (PDR-Bridge) that performs safeguarded local refinement and outputs transferable primal, projection-dual, sensitivity, and health tokens. Across IEEE 14-, 30-, 57-, 118-, and 300-bus systems, the benchmark contains 26.3 million solver-generated perturbation-response pairs; in the 300-bus transfer setting, pretraining uses only the 21.04 million nominal pairs from the 14-, 30-, 57-, and 118-bus source topologies. Across three runs on unseen IEEE 300-bus transfer with 10% labels, PhysBridge obtains a 2.21±0.05% OPF cost gap, 96.3±0.2% feasibility, 1.84±0.05 $/MWh LMP MAE, and 91.6±0.3% contingency/fault event macro F1. Relative to a matched rollout-only model trained on the same source corpus without response targets or bridge refinement, these values correspond to a 50.8% lower OPF gap and a 7.4 percentage-point higher event F1. The bridge requires 12.4 ms per 300-bus forward pass and activates fallback on 4.6% of stressed windows. Comparisons with supervised-only OPF specialists that do not receive matched solver-generated pretraining are interpreted as task-relevant performance under the proposed pretraining regime, not as architecture-only superiority.
SN - 1932-6203
ER -