【作者】 胡志海；

【导师】 刘圣军； 刘新儒；

【作者基本信息】 中南大学 ， 应用统计， 2023， 硕士

【摘要】 近年来长江中下游流域洪涝灾害频发,给群众的生命财产安全带来重大损失,而极端降水是引发洪涝灾害的直接原因,因此研究长江中下游流域的极端降水对于洪涝灾害的防治具有重要的意义。本文主要从特征分析和预测分析两个方面对长江中下游流域极端降水进行分析,在特征分析方面,本文计算出了能够描述极端降水特征的极端降水指数,并运用趋势分析、突变检验、小波分析、概率分布拟合分析过去几十年间8种极端降水指数的特征,发现极端降水发生的频率和强度均在增加,此外通过概率分布的拟合可以计算给定阈值的概率,这在风险评估中具有重要作用。在预测分析方面,本文利用对降水有重要影响的10个环流因子来预测4个极端降水指数,通过加入正则化的线性回归方法及其集成方法进行拟合,发现L2正则化的拟合效果最好;在机器学习方法上,本文运用了支持向量机、决策树、随机森林以及集成学习模型,发现支持向量机和随机森林模型表现较好。在时间序列模型上,本文用ARIMA模型对月尺度降水数据进行预测,加入季节要素后,模型的效果得到提升。在深度学习方法上,本文使用了LSTM和GRU模型对日数据进行建模,在构造样本时,采用过去时期的数据对未来数据进行预测,分别用两种模型的最优方法对未来3年的日降水量进行预测,并用预测值计算了未来3年的极端降水指数,发现在未来3年极端降水的总量和降水日数都比往年平均水平高。在模型预测时,如果预测步长较长,后期的预测值会趋于平稳,需要从样本构造、网络设计等方面进行进一步调整和优化。图23幅,表19个,参考文献100篇更多还原

【Abstract】 Extreme precipitation is the direct cause of flooding,so it is important to study the extreme precipitation in the middle and lower reaches of the Yangtze River basin for flood prevention and control.In this thesis,extreme precipitation indices describing the characteristics of extreme precipitation are calculated,and the characteristics of eight extreme precipitation indices over the past decades are analyzed by trend analysis,abrupt change test,wavelet analysis,and probability distribution fitting,and it is found that the frequency and intensity of extreme precipitation are increasing.In addition,the probability of a given threshold can be calculated by fitting the probability distribution,which has an important role in risk assessment.In terms of prediction analysis,10 circulation factors that have important effects on precipitation are used to predict 4 extreme precipitation indices,which are fitted by linear regression methods with regularization and their integrated methods.models performed better.In terms of time series model,this paper uses ARIMA model to predict monthly scale precipitation data,and the effect of the model is improved by adding seasonal elements.In terms of deep learning methods,this paper used LSTM and GRU models to model daily data,and in constructing the samples,data from past periods were used to predict future data,and the optimal methods of the two models were used to predict daily precipitation for the next three years,respectively,and the extreme precipitation index for the next three years was calculated using the predicted values,and it was found that the total amount of extreme precipitation and the number of precipitation days in the next three years were higher than the average level is higher.In the model prediction,if the prediction step is long,the prediction value will tend to be smooth in the later period,and further adjustment and optimization are needed from sample construction and network design.更多还原