关键词: 废钢精细配比推荐, 大语言模型, 知识图谱, 链式推理, 转炉炼钢

Abstract: Aiming at the problem that the converter scrap proportioning heavily relies on expert experience and historical smelting knowledge is difficult to reuse efficiently,making it challenging to adapt to dynamically changing smelting conditions,a Large Language Model(LLM) enhanced Knowledge Graph(KG) chain reasoning method for refined converter scrap proportioning recommendation was proposed.Based on logical chain decomposition,a query decomposition and reasoning method and neighborhood retrieval was proposed.By defining four types of first-order logical queries and designing a query decomposition method,complex multi-hop reasoning scrap proportioning queries were decomposed into sequences of single-hop reasoning sub-queries to improve KG reasoning accuracy and stability.Furthermore,a neighborhood retrieval-based reasoning strategy was designed to enhance KG reasoning efficiency by prioritizing the traversal of subgraphs in the neighborhood of the queried entities.Then,an LLM-enhanced KG chain reasoning architecture was constructed.By constructing structured prompt templates,the relevant subgraphs and sub-logical query sequences were converted into contextual prompts and chained problem prompts respectively,which guided LLM to perform step-by-step reasoning on the KG,thus the sub-logical queries were solved sequentially,and the intermediate results were integrated to generate scrap proportioning recommendation schemes that met practical smelting requirements.Finally,the proposed method was validated using a converter smelting case from a steel plant,and the results demonstrated that the proposed KG-LLM architecture effectively improved the decision-making accuracy for scrap steel proportioning,with performance further enhanced particularly after introducing the chain-of-thought mechanism,confirming the effectiveness and superiority of the proposed method.

Key words: scrap refined proportion recommendation, large language model, knowledge graph, chain reasoning, converter steelmaking