The integration of renewable energy sources and the increasing complexity of distribution networks have created not only accuracy challenges but also substantial computational demands for real-time line loss analysis across large numbers of feeders. To address these issues, this paper proposes the line loss graph transformer (L2GT), which combines graph convolutional networks (GCNs) and transformer-based multi-head attention to jointly model local topology-aware interactions and long-range dependencies on graph-structured power-grid data. From a computational perspective, the proposed framework leverages sparse graph operations and GPU-parallel attention computation to support high-throughput inference for operational deployment. Experimental results on real distribution network data and large-scale synthetic topologies demonstrate that L2GT significantly outperforms existing methods in prediction accuracy and robustness, while maintaining millisecond-level inference latency and scalable performance up to 5000 nodes. These results indicate that L2GT is not only an accurate line loss prediction model but also a high-performance computing-oriented framework for real-time monitoring and analysis in modern distribution networks.
Zhu et al. (Mon,) studied this question.