Proceedings of the PowerUp Conference

Distribution utilities increasingly rely on data-driven methods for state estimation and forecasting as large-scale renewable energy integration causes rapid state changes, yet most learning-based approaches remain tied to specific feeder topologies and require retraining when networks are reconfig- ured. In practice, retraining models for every feeder and opera- tional configuration is infeasible due to limited data availability, modeling effort, and computational costs. This paper presents a physics-aware graph learning framework for three-phase un- balanced distribution networks that enables zero-shot transfer across unseen feeder reconfigurations, with potential extension to feeders sharing similar electrical and topological statistics. By combining scalar-weight graph convolutions grounded in power flow physics with adaptive graph pooling, the proposed model learns topology-invariant representations while accommodating variable network sizes. Case studies on the IEEE 123-bus unbalanced distribution feeder using AMI-based measurements demonstrate accurate state estimation and short-term forecasting under unseen reconfigurations, highlighting the promise of the approach for reconfigurable distribution systems.