【作者】 黄鑫；

【导师】 于达仁； 王华宁；

【作者基本信息】 哈尔滨工业大学 ， 动力工程及工程热物理， 2010， 博士

【摘要】 太阳耀斑作为最剧烈的太阳活动形式之一,它的爆发直接影响空间和地球的环境,进而影响人类的生产和生活。因此,耀斑预报的研究具有重大的实用价值。随着观测手段和设备的不断改进,人们能够得到大量的太阳观测数据。如何从海量的观测数据中获得知识、建立预报模型成为太阳物理研究中的一个关键的问题。本文基于太阳活动区光球磁场观测和耀斑观测数据,研究了自动知识发现和预报模型建立的方法。具体从以下几个方面进行了探索:(1)在建模理论的指导下,确定模型的动态特性,给出需要引入的观测序列的长度,并给出基于滑动窗方法的耀斑预报模型。相对于现阶段使用的耀斑模型,本文给出的耀斑预报模型能够反映活动区观测的动态信息对耀斑的影响,是一个动态模型。考虑到活动区的演化信息,耀斑预报的建模问题可以看成是机器学习中的贯序监督学习问题。本文利用滑动窗方法,将贯序监督学习问题转化为一般的监督学习问题加以解决,并验证了预报因子序列在预报模型中的重要作用。(2)构造光球磁场序列的多尺度预报因子。为了更加全面地反映活动区光球磁场序列的演化特性及其对耀斑预报的影响,本文基于极大重叠离散小波变换和序列特征提取方法构造了耀斑的多尺度预报因子。相对于现有的从活动区磁图中提取出的空间多尺度预报因子,本文第一次构造出反映活动区时间演化特性的多尺度预报因子。利用信息论中信息增益率的概念,定量地刻画了多尺度预报因子的耀斑预报能力。并给出具有较强预报能力的多尺度预报因子的物理解释。选择预报能力强的多尺度预报因子,建立了基于多尺度预报因子的太阳耀斑预报模型,验证了多尺度预报因子的重要作用。(3)建立耀斑预报的不确定性规则模型。由于对耀斑的物理本质认识得不够深入,现阶段提取的预报因子与预报模型仅存在一定的概率依赖关系。为了更好地表达这种关系,利用贝叶斯网方法从观测数据中建立了耀斑预报的不确定性规则模型,并给出所建模型的物理解释。相对于现有的活动区观测特征与耀斑间的确定性关系,本文首次从数据中学习到活动区观测与耀斑间的不确定性关系。该方法不仅可以用来建立耀斑预报模型,更重要的是它还可以用于观测数据中的知识发现。面对日益增多的观测变量和观测数据量,这无疑是一个有重要意义的研究方向。(4)提出预报因子组的概念,基于多个预报因子组建立了耀斑预报的多模型融合模型。相对于现有预报模型以单个预报因子为基本的输入单元,本文首次提出预报因子组的概念,解释了形成预报因子组的合理性,并将预报因子组作为耀斑预报模型的基本输入单元。由于预报因子组的不唯一性,为每个预报因子组建立一个耀斑预报的基模型,然后将这些模型的结果组合起来,形成耀斑预报的多模型融合模型。更多还原

【Abstract】 Solar ?are is one of the most severe solar activities. It in?uences the space weatherand some activities on the Earth, so it is valuable to predict the level of solar ?ares.With the development of the observational instruments, large amounts of data is obtained.One of the most important scientific problems is how to extract knowledge and buildprediction model from the data. This problem is discussed and the main contributions ofthis dissertation are listed as follows:(1) Under the guidance of the modeling method, the dynamic characteristics ofthe prediction model are determined, and the prediction model with the sliding windowmethod is built. Comparing with the current prediction models, the proposed model,which can re?ect the evolutionary information of the photospheric magnetic field in theactive regions, is a dynamic model. Taking into account the evolutionary information ofactive regions, building a ?are prediction model can be viewed as a sequential supervisedlearning problem in machine learning. Here, the sequential supervised learning problemis transformed into the standard supervised learning problem, and the importance of thesequence of predictors is validated.(2) Multiscale predictors of photospheric magnetic field are proposed. In order tofully describe the in?uence of the evolution of photospheric magnetic field in active re-gions on the eruption of solar ?ares, multiscale predictors are constructed using maximumoverlap discrete wavelet transform and sequential feature extraction method. Compar-ing with the existing multiscale predictors extracted from photospheric magnetograms,the proposed multiscale predictors re?ect the evolutionary characteristics of photosphericmagnetic field. The predictability of the proposed multiscale predictors is quantitativelyestimated by information gain ratio, and the physical explanation of these predictors isgiven. Using these predictors, the solar ?are prediction model is built, and the effective-ness of these predictors is validated.(3) The uncertainty prediction model of solar ?ares is established. Because of thelimitation on the physical understanding of solar ?ares, the predictors probabilisticallyrelate to the eruption of ?ares. Bayesian network learned from the observational data isused to express these relationships, and the physical interpretation of this model is given. Comparing with existing models with the deterministic relationships, the uncertainty pre-diction model of solar ?ares is firstly developed. This model not only can be used toforecast the eruption of ?ares, but also can be used for knowledge discovery from theobservational data. For the increasing quantity of the observed data, this will become animportant research direction.(4) The concept of predictor teams is proposed, and the multiple prediction modelsbuilt by predictor teams are fused. The predictor team is firstly proposed and the reason-ability of the predictor team is explained. The base prediction models of solar ?ares arebuilt using predictor teams, and then these base models are fused to generate a compre-hensive prediction model.更多还原