Knowledge graph enhanced semantic parsing for distribution networks using graph convolutional encoding and topology-constrained decoding

Text-to-Cypher semantic parsing is a promising interface for distribution-network knowledge graphs, yet most existing parsers flatten topology into text and leave the decoder only weakly constrained. This often produces executable-looking queries with broken relation chains, incorrect entity grounding, or topological hallucinations in multi-hop reasoning. We propose KGE-SP, a topology-aware semantic parsing framework that combines a RoBERTa intent encoder, a residual graph convolutional topology encoder, cross-modal fusion, and a masked pointer decoder. During autoregressive generation, the decoder can select only those entities and relations that remain valid under the current subgraph state, which keeps the query trajectory consistent with physical connectivity. We also construct DN-SP, a benchmark containing 18 542 natural-language-to-Cypher pairs derived from SCADA alarm logs and GIS asset records. On DN-SP, KGE-SP achieves 99.1% valid Cypher rate, 82.4% exact match, 85.1% execution accuracy, and 87.6% execution F1, outperforming sequence-generation, retrieval-augmented, and large-language-model baselines. Cross-regional zero-shot evaluation further shows improved robustness under feeder reconfiguration while keeping average inference latency below 115 ms. These results demonstrate that explicit topology injection is an effective route toward reliable graph-grounded semantic parsing in distribution automation.