近年来，突然事件日趋复杂且高度不确定，应急决策分析面临多重难题：多决策任务并发且相互影响、缺乏有效的模型来准确地预测未来、预测准确性不足导致难以支持合理有效的策略选择。面向复杂突发事件应急决策的困难和挑战，本文探索了一种新的决策分析方法。具体地，本文聚焦复杂突发事件的问题建模、仿真预测以及策略评估与选择三项核心行动，分别提出了基于多模型耦合的集成建模方法、动态探索性的仿真方法以及鲁棒决策评估模型与优化方法，并进行应用与验证，最后将三种方法融合至统一的决策分析框架内，形成了复杂突发事件的动态鲁棒决策分析范式。 针对复杂突发事件存在的多个并发且互相关联的决策任务，本文提出了基于多模型耦合的集成建模方法。该方法包括分布式异构模型的深度共享与集成架构、模型检索与组合机制以及组合模型评价与动态配置方法组成，可以为聚集、发现、耦合不同模型，建立集成建模系统提供全流程方法指导。本文以危化品泄漏背景下的区域紧急疏散事件为例，进行了方法验证，结果证明该方法能够有效集成跨学科的、分布式的、异构的模型，实现复杂决策问题的系统化、整体化建模。 针对不确定性导致的突发事件仿真不准确、不可靠的难题，本文提出了情景计算方法，通过动态“数据+模型”混合建模和多样性计算实验，使计算系统建立了持续探索、重构、改进系统模型及参数的能力。本文将该方法应用至交通疏散领域，对真实的机动车跟驰行为进行了预测，并与其他仿真方法进行对比。结果表明，情景计算方法能更有效地提升预测的准确性，同时帮助计算系统灵敏地适应真实系统的不确定变化，并探索真实系统的行为规律。 针对深度不确定条件下模型预测结果准确性、可靠性不足的问题，本文提出了一种鲁棒决策评估模型与策略优化方法。鲁棒决策模型引入遗憾值度量策略的性能，能从大量不可靠的预测结果中识别遗憾值较低且对未来不敏感的鲁棒策略；策略优化方法能够对多样性的预测结果进行聚类分析，从而识别鲁棒策略潜在的脆弱性与机遇性，帮助决策者动态地改进策略。该决策模型与优化方法有望解决具有不可完全预测与不可控特征的突发事件策略评估、选择与优化问题。 本论文成果形成了创新性、成体系的决策分析理论与方法，有望为具有高度复杂性和动态不确定性的突发事件应急决策提供理论指导与技术支持。

In recent decades, emergencies are becoming increasingly complex and uncertain. Decision analysis for emergency management often encounters with multiple significant challenges: 1) there are multiple interdependent decision problems in one case, 2) decision makers cannot specify appropriate models or parameters to accurately predict the future, and 3) predictive information produced by computer models are too unreliable to support the selection of an effective contingency plan. To date, no general approaches have been available for decision analysis in complex emergencies. This study aims to propose such a decision analytic approach. This study chooses three key issues that are “how to model complex problems”, “how to predict uncertain future” and “how to find a satisfied strategy”, and respectively provides a novel solution. Specifically, this study developed an integrated modelling method based on multi-models coupling, a dynamic exploratory simulation method for the systems with deep uncertainty, and a robust decision model along with a strategy optimization mechanism. This paper has validated these methods with case applications in regional emergency evacuation. This study incorporates the methods into a unified decision analysis procedures to build a general dynamic robust decision analytic approach for management of complex emergencies. There are always many problems coexisting and interacting with each other in a complex emergency event. This study developed an integrated modelling pathway that can combine different computer models to model complex problems in a holistic manner. This pathway includes a model sharing and integration architecture, a model discovery and combination mechanism, a validity evaluation method and a dynamical configuration algorithm of composite model. The pathway provides a systematic and operable guideline for decision makers to aggregate, search and couple originally dispersed, heterogeneous and transdisciplinary models to construct an integrated modelling system. This study demonstrates the pathway with a case application in which the whole planning process of a regional emergency evacuation caused by hazardous material leaking was successfully modelled by different models. The results show that the integrated modelling pathway can integrate distributed heterogeneous models to achieve a holistic modelling of complex decision problems. Due to the existence of deep uncertainty, model-based simulation cannot provide the future accurately. To cope with uncertainty, this study putted forward a dynamic exploratory simulation method named “Scenario Computing”. Scenario computing is to 1) hybridize available data and models to build diverse initial conditions, 2) use these initial conditions to generate a wide diversity of plausible futures, and 3) explore potential patterns or mechanisms of real system with analysis and visualization of massive futures and hence to improvise or find new plausible model(s) for adjusting computing system. This study applied the proposed method in traffic evacuation field to predict car-following behavior that is filled with human uncertainties. By comparing to other simulation methods, it is found that scenario computing can better improve the prediction accuracy and help computing system adapt to uncertain changes of real system sensitively and explore potential evolution laws of real system. As model-based predictions of the future are highly unreliable in complex emergencies, it will be problematic to select a strategy with optimal outcomes. This study proposed a robust decision model and a strategy optimization method to support alternative selection. This robust model employs regret value to measure the strategy performance and aims to select the robust strategy with relatively small regret compared to the alternatives and low sensitivity to the uncertain future. The proposed strategy optimization method is to dynamically identify potential vulnerabilities and opportunities of robust strategy via the clustering analysis of diverse predicted results. It can help decisionmakers improve the robust strategy overtime. The robust decision model along with strategy optimization method may have capability to support the alternative evaluation, selection and optimization of emergency decision characterized by both uncontrollable and unpredictable. In conclusion, this study presents a novel, systematic decision-analytic approach to complex emergencies. It could find a feasible pathway for emergency decision analysis under the conditions of high complexities, dynamics and uncertainties and provide operable guideline for handling practical disaster decision problems.