Mitigating cross-class semantic interference: a hierarchical class-constrained pattern matching method for distribution network CIM modelling

Distribution network business systems require consistent CIM modelling across equipment ledgers, monitoring, control and dispatching applications, but heterogeneous vendor ledgers often contain inconsistent field names, granularities and device categories. This paper proposes a hierarchical class-constrained schema matching method for mapping distribution-network business fields to CIM attributes. The method first retrieves candidate CIM device classes from the co-occurrence context of source-table fields, then performs attribute matching within the candidate classes and their inheritance paths, reducing interference from sibling classes, inherited attributes and generic fields. Two implementations are provided: an LLM reranking route for accuracy-prioritized scenarios and a graph reranking route for local low-resource deployment. On 17 desensitized vendor source tables with 194 labelled fields and 13 CIM classes, Two-Stage-LLM achieves an F1 score of 0.9444, outperforming Flat-LLM by 0.1172, while Two-Stage-Graph improves Flat-Graph from 0.4100 to 0.7304 without external LLM calls.