【作者】 徐磊；

【导师】 李向阳；

【作者基本信息】 哈尔滨工业大学 ， 管理科学与工程， 2013， 博士

【摘要】 由于突发事件具有事发突然、事态复杂、高度不确定性、涉及范围广等特点，使其应急决策过程存在决策时间短、应急预案少、决策复杂度高等问题。明显地，传统的基于预案的应急决策方法无法很好地满足重大突发事件应急决策的需求。因此，如何基于突发事件发展过程中所获取的信息，对其进行充分有效的分析，用以研究和发现突发事件的演化趋势和规律，为应急方案的快速准确的建立提供技术支持，是当前突发事件应急决策管理领域面临的关键问题。在信息化理论和技术飞速发展的新时期，将人工智能理论和相关技术与突发事件应急决策支持方法相融合，在对事件数据进行处理和分析的基础之上，预测突发事件的演化趋势，评估突发事件的演化状态，是为突发事件应急决策提供有效支持的重要方法之一。贝叶斯网络作为人工智能研究领域的一个重要分支，结合了图论和统计学方面的知识，能够以一种自然的方法表示因果信息，具有丰富的概率表达能力、不确定性问题处理能力、多源信息表达与融合能力。因此，将贝叶斯网络与突发事件应急决策相融合，研究基于贝叶斯网络的突发事件应急决策信息分析方法具有十分重要的理论意义和现实意义。本文对国内外相关研究成果进行了梳理和总结，在此基础之上，以突发事件应急管理理论、系统动力学理论、贝叶斯网络理论、案例推理理论等作为理论基础，以相关技术为支持，综合运用定性分析和定量分析等研究方法，对基于贝叶斯网络的突发事件应急决策信息分析方法进行了深入的研究。首先，以突发事件演化系统结构作为对象，研究了面向突发事件应急决策支持的贝叶斯网络结构学习方法。针对突发事件应急决策所面临的先验知识和数据信息的不完备性、事件发展的不确定性以及数据量大等问题，通过研究分析已有的贝叶斯网络结构建模方法，融合K2和MCMC贝叶斯网络结构学习方法的优势，基于模型平均的思想，提出了基于模型平均的贝叶斯网络结构学习改进方法，以及可以有效提高贝叶斯网络计算效率的拓扑结构并联优化方法。其次，基于系统动力学理论以及贝叶斯网络理论和方法，研究了基于贝叶斯网络的突发事件灾情评估方法。基于系统动力学对突发事件演化的关键参数进行了分析，得到了不同子系统关键参数之间的因果反馈关系，进而，针对贝叶斯网络参数会随建模环境不断变化的问题，考虑到突发事件应急决策所需的实时更新能力，基于最大似然估计、最大后验概率和最大期望算法，提出了贝叶斯网络参数学习的改进方法——实时增量式最大期望参数学习方法，并且通过输电网络灾情评估仿真案例对其进行了验证。再次，针对突发事件应急决策的复杂性和时间敏感性，提出了基于贝叶斯网络的应急决策支持案例适配方法。将历史贝叶斯网络模型作为案例存储于案例库中，通过案例检索和匹配、案例修正等方法构建新的贝叶斯网络模型，以达到复用贝叶斯网络历史模型的目的。该方法没有庞大的搜索空间，也不需要样本数据，只需要提前收集历史案例模型，与传统贝叶斯网络建模方法相比具有更高的建模效率，可以作为突发事件应急决策支持贝叶斯网络建模方法的一个重要辅助方法。最后，对贝叶斯网络模型在洪涝灾害应急决策信息分析中的应用进行了研究。基于本文提出的面向突发事件应急决策信息分析的贝叶斯网络方法，对洪涝灾害进行了分析，从洪涝灾害演化和洪涝灾害应急决策支持两个方面进行了分析和研究，建立了相应的贝叶斯网络模型，从而为洪涝灾害应急决策提供参考依据。进而，验证了本文提出的基于贝叶斯网络的突发事件应急决策信息分析方法的有效性。更多还原

【Abstract】 Emergent events are bursty, complex, extremly uncertain, and widely involved,which makes it a highly complicated problem and needs the decision makingprocess to be done within a short time with fewer pre-arranged plannings.Obviously, the traditional decision-making methods with pre-arranged plannings cannot well meet the needs of the emergency management.Therefore, how to fast andeffectively analyse the data and evolution laws of emergent events in order tosupport the decision making with referential information has become a key issue inthe field of emergent events decision-making and management.Under the circumstance of the industrial informationization, emergent eventsdecision-making support methods with artificial intelligence technology can forecastand assess the emergency tendency based on effectively analysis of emergent eventsdata, which shows a new way to support emergent events decision making. As animportant branch of artificial intelligence research, Bayesian networks is combinedwith the knowledge of graph theory and statistics, and can manage uncertaintyproblems, express and fuse multi-source information. Therefore, applying Bayesiannetworks into emergent events decision making, researching the emergent eventsdecision-making information analysis method based on Bayesian networks hasnotable theoretical and practical significance.The domestic and foreign research achievements are analyzed and summarized.Then, based on emergency management theory, system dynamics theory, Bayesiannetwork theory, case reasoning theory and accordingly techniques, the emergentevents decision-making information analysis method based on Bayesian networks isdeeply researched by qualitative and quantitative analysis.First, taking the system structure of emergent events evolution as the researchobject, the Bayesian network topology learning method for supporting emergentdecision making of emergent events is studied. In order to adapt to incompleteinformation, uncertain development and massive data, based on model average, animproved Bayesian network topology learning method which combined bothadvantages of K2and MCMC is proposed. A Bayesian Network topologyoptimization method is also proposed to improve the system efficiency.Second, based on system dynamics and Bayesian network theories, theemergency assessment method based on Bayesian network is studied. By analyzingthe key parameters of emergent events evolution, the causal feedbacks of keyparameters within different subsystems are obtained. And then, based on maximumlikelihood estimate, maximum a posteriori estimation and expectation maximization algorithm, an improved Bayesian network parameter learning method is proposed inorder to adapt to the changing modeling circumstance and the need of real-timeupdate ability. An example application to a power network disaster assessment ispresented to illustrate and test the proposed method.Third, in order to adapt to the complexity and time sensitive of the emergentevents decision making, a case adaptation method based on Bayesian network foremergent events decision making is proposed. The historical Bayesian networkmodels are stored in a case database, a new Bayesian network model can beobtained by case retrieve and case revision, which can enhance the reuse ofhistorical models. The proposed method do not have a huge search space or needsample data. Compared with the traditional Bayesian network modeling method, thismethod is more effcient and can be an important assistant to the Bayesian networkmodeling method for emergent events decision-making support model.Finally, an application of Bayesian network for emergency decision making inthe flood disaster is illustrated. With the analysis of the flood evolution andemergency decision making for flood, the Bayesian network models are built basedon the proposed Bayesian network method for emergent events decision makinginformation analysis, which shows the ability to support the flood disasteremergency decision making with referential information. Therefore, the proposedemergent events decision making information analysis method based on Bayesiannetwork is verified.更多还原