Knowledge graph-enhanced carbon impact assessment of construction and demolition waste with large language model-based decision support

The construction and demolition waste (C&DW) sector generates substantial carbon emissions across sorting, transportation, and disposal phases. However, existing assessment approaches often generate static numerical results that offer limited traceability, regulatory interpretation, and decision support. This paper proposes an integrated framework combining a domain-specific knowledge graph (KG) with a large language model for C&DW carbon management. A C&DW Carbon KG is constructed in Neo4j to connect waste categories, treatment methods, facility information, source-referenced emission factors, regulatory constraints and calculated carbon results. A retrieval-augmented generation (RAG) mechanism connects the KG with GPT-4o to support evidence-grounded explanations. In a case study of a 12-storey residential building in Shenzhen, transportation was the dominant emission phase, and sensitivity and scenario analyses showed that the carbon benefit of recycling depended on facility distance and recycling rate. In benchmark testing, full KG-RAG achieved 0.918 overall accuracy versus 0.410 for document RAG and 0.159 for an LLM-only baseline, abstaining on all out-of-scope queries with no fabricated values. This paper contributes an application-oriented approach that integrates carbon accounting, knowledge organization, and regulation-aware explanation to support traceable and explainable decision-making in low-carbon C&DW management.