基于多模型融合与检索增强生成的危险化学品事故智能应急处置决策支持系统设计

Design of an Intelligent Emergency Response Decision Support System for Hazardous Chemical Accidents Based on Multi-Model Fusion and Retrieval Enhancement Generation

摘要: 针对危险化学品事故具有突发性强、演化过程复杂以及现场应急决策信息需求高等特点，传统应急处置方式存在数据整合效率低、事故后果定量分析不足以及辅助决策智能化水平有限等问题，提出一种融合事故后果预测模型、专业知识检索和端侧大语言模型的危险化学品事故智能应急处置决策方法。该方法以危险化学品基础数据为支撑，通过构建专业知识库实现事故案例、法规标准和现场处置资料的结构化管理；结合检索增强生成(Retrieval-Augmented Generation, RAG)技术，将专业知识检索结果与人工智能模型推理过程进行融合，提高智能生成结果的专业性和可追溯性。同时，针对危险化学品泄漏、火灾、爆炸以及有毒气体扩散等典型事故场景，集成泄漏源项计算、池火灾、沸腾液体扩展蒸气爆炸(BLEVE)、蒸气云爆炸(UVCE)和有毒气体扩散等事故后果预测模型，实现事故影响范围的定量分析。在此基础上，开发危险化学品事故应急处置决策支持系统，实现事故信息管理、风险后果计算、消防救援分析以及智能处置方案生成等功能。系统采用桌面端与移动端协同架构，并支持完全离线运行，可满足化工生产现场和应急救援现场网络受限条件下的快速决策需求。研究结果表明，该系统能够实现危险化学品事故信息快速获取、风险区域辅助判定以及应急处置方案智能生成，为事故现场科学决策提供技术支持。

Abstract: To address the characteristics of hazardous chemical accidents, such as strong suddenness, complex evolution processes, and high demands for on-site emergency decision-making information, traditional emergency response approaches suffer from issues including low data integration efficiency, insufficient quantitative analysis of accident consequences, and limited intelligence in auxiliary decision-making. This paper proposes an intelligent emergency response decision-making method for hazardous chemical accidents that integrates accident consequence prediction models, professional knowledge retrieval, and on-device large language models. Supported by fundamental hazardous chemical data, the method constructs a professional knowledge base to enable structured management of accident cases, regulatory standards, and on-site disposal documents. By incorporating Retrieval-Augmented Generation (RAG) technology, it integrates professional knowledge retrieval results with AI model inference processes, thereby enhancing the professionalism and traceability of intelligently generated outputs. Furthermore, for typical accident scenarios including hazardous chemical leaks, fires, explosions, and toxic gas dispersion, the method integrates consequence prediction models such as leak source term calculation, pool fires, boiling liquid expanding vapor explosion (BLEVE), unconfined vapor cloud explosion (UVCE), and toxic gas dispersion, enabling quantitative analysis of the accident impact range. On this basis, a decision support system for hazardous chemical accident emergency response is developed, featuring functions such as accident information management, risk consequence calculation, firefighting and rescue analysis, and intelligent response plan generation. The system adopts a collaborative desktop-mobile architecture and supports fully offline operation, meeting the rapid decision-making needs under network-constrained conditions at chemical production sites and emergency rescue scenes. The research results demonstrate that the system enables rapid acquisition of hazardous chemical accident information, auxiliary determination of risk zones, and intelligent generation of emergency response plans, providing technical support for scientific decision-making at accident scenes.