Event logic graph-augmented generation for interpretable urban waterlogging emergency decision-making

Urban waterlogging emergency assessment requires credible causal tracing, factual grounding, and actionable response planning. However, directly applying mainstream Retrieval-Augmented Generation (RAG) methods to this high-reliability scenario still faces challenges, because similarity-oriented retrieval may retrieve semantically relevant but causally incomplete fragments, providing insufficient support for traceable emergency decision-making. To address this problem, this paper proposes an Event Logic Graph-Augmented Generation (ELGAG) framework. ELGAG constructs an urban waterlogging Event Logic Graph (ELG) to organize historical cases, causal factors, response actions, responsible entities, and emergency resources into event-centered evidence chains. Based on the ELG, a multi-constraint matching and causal aggregation retrieval (MCM-CAR) algorithm is designed to match current emergency queries with historical events and aggregate structured cause–event–response evidence for report generation. A Large Language Model (LLM) then generates emergency assessment reports under domain knowledge-enhanced prompt constraints, while a validation-gated feedback mechanism supports controlled knowledge evolution. Experiments on Guangzhou urban waterlogging data from 2010 to 2019 show that ELGAG achieves a Weighted Overall Quality Score (WOQS) of 0.775 ± 0.001 over five repeated runs and obtains the highest overall performance among vector-based and graph-enhanced RAG baselines under the same evaluation protocol. Ablation, robustness, sensitivity, feedback-update, and LLM-backbone analyses further indicate that the performance gain mainly comes from event-logic representation, MCM-CAR retrieval, and evidence-grounded generation. This study provides an interpretable and verifiable paradigm for applying LLMs to safety-critical urban waterlogging emergency decision support.