极值统计理论的进展及其在气候变化研究中的应用

着重论述极值统计分布在极端天气气候事件和重大工程设计中的重要意义,综述该领域国内外研究进展。例如,基于超门限峰值法（POT）的广义帕累托分布（GPD)和基于单元极大值法（BM）的广义极值分布（GEV）及其参数间的理论关系;采用极值分布模型与多状态一阶Markov链相结合构建降尺度模型模拟局地极端降水事件,推算一定重现期的极端降水量的分位数;探讨极值分布模型分位数估计误差问题,多维极值分布理论及其应用等问题。     

南京过去100年极端日降水量模拟研究

在南京过去100年日降水资料的基础上,利用极值理论中的区组模型和阈值模型分析了极端日降水分布特征.首先通过广义极值(GEV)模型模拟了日降水的年极值序列(AMDR),用极大似然估计(MLE)方法计算了模型的参数,并借助轮廓似然函数估计出参数的精确误差区间,同时采用4种较直观的诊断图形对模型的合理性进行全面评估,结果表明Frechet是区组模型中最适合描述极端日降水分布特征的函数.其次,将日降水序列分3种情景构建极值分布的阈值模型(GPD),考察了观测数据的规模对应用该模型的限制,重点讨论了如何针对给定观测样本选择合适的阈值收集极值信息.分析结果认为,长度不小于50年的气候序列,采用24 mm的日降水量作为临界阈值均能进行GPD分析.该阈值处于年降水序列第91个百分位附近,即对目前长度为50年左右的日观测资料,第91个百分位点以上的数据基本能满足GPD研究的需要.另外,根据GEV和GPD对未来极端降水重现水平的推断情况,GPD预测值的置信区间要比GEV的窄,极值推断的不确定性相对也较小,更适合用于研究中国目前规模不大的气候资料.最后,对GPD模型的形状参数和尺度参数进行变换,分别引入描述线性变化的动态变量,分析降水序列中潜在的变异行为对极值理论应用的影响.这种变异包括降水序列中长期的均值变化及百分位变化,从模拟结果看,暂未发现资料变异行为对极值分析产生显著于扰.     

利用广义帕雷托分布拟合中国东部日极端降水的试验

引进广义帕雷托分布拟合我国东部地区78个测站夏季(5～9月)逐日极端降水量.结果表明,不同门限值条件下的逐日降水量所拟合的降水极值概率分布均符合广义帕雷托分布,与其它极值分布如广义极值(下称GEV)分布模式相比,以GPD模式为最优.根据现代气候条件,分别计算了50年一遇和100年一遇的极端降水量分位数并分析其空间分布特征,两者基本一致,总体上都呈现出由东南向西北方减小的趋势,且南北差异较大,南方的极端降水量值可能达到北方地区的两倍以上.此外,资料年份越长,拟合效果越好.     

广义极值分布理论在重现期计算的应用

在气候统计学上,常用Weibull、Gumbel、Frechet统计分布函数对极端气候要素的分布进行拟合,广义极值分布理论综合了以上三种极值分布模型,在气候分析中得到了广泛应用。以南昌市年汛期日最大降水量为例,利用广义极值分布理论对其分布进行拟合,并对重现值及其置信区间进行计算,为气候要素极值的统计分析提供了一种新的手段。     

四川盆地短历时强降水极值分布的研究

运用广义帕雷托分布(GPD)和广义极值分布(GEV),借助于L-矩的参数估计方法,对四川盆地12站的小时极端降水量进行拟合,并对两种模型的拟合效果进行比较。运用Hill图,结合统计量D*来确定GPD的最佳门限值是合适的,选出的样本是独立的。各站的小时极端降水概率分布均符合GPD和GEV,但GPD模型的拟合精度要优于GEV模型。利用两种模型推算出各站给定重现期的最大小时降水量,其中泸州50 a一遇和100 a一遇的降水极值分位数都超过了100 mm,除了遂宁站外,两种模型估计出的极值分位数的相对误差基本都在10%以下。通过分析,GPD推算的结果更加可靠。     

中国格点化日降水极值统计模型及阈值的选取

采用年最大值法（AM）及超阈值峰量法（POT）分别构建基于0.5°×0.5°网格的全国地面日降水极值序列,建立基于广义极值分布（GEV）和广义帕累托分布（GPD）的降水极值统计模型,通过K-S检验评估模型拟合效果,研究全国日降水极值的统计规律及其空间分布特征,提出适用于不同地区极端日降水的极值分布模型与阈值选取标准,结果表明：(1)POT序列比AM序列更符合降水极值序列的要求;(2)为便于比较并提高模型拟合效果,POT序列的阈值由百分位数法确定效果较好;(3)阈值方案优选结果在空间分布上与中国干湿区域的划分有很好的相关性,在湿润地区宜将第90～94百分位数作为阈值,在半湿润和半干旱地区宜将第94～97百分位数作为阈值,在干旱地区则使用第97～99百分位数较为合适。     

长江三峡地区宜昌、巴东短历时极值降水特征分析

利用长江三峡库首宜昌站及库区巴东站1955—2008年分钟降水强度资料,采用广义极值分布和线性矩参数估计方法,拟合两站7个短历时(60min以内)年最大降水量概率分布,推断各历时有关重现期降水极值,计算各历时暴雨频次及年最大降水量气候倾向率,分析各历时降水广义极值分布的参数随时间变化规律。结果表明:宜昌、巴东两站7个短历时年最大降水量采用广义极值分布拟合,其效果较好;两站短历时降水平均值趋势变化不明显,而不同百分位数降水量变化趋势差异较大,其中中位数的降水量呈下降趋势,较高百分位数的降水量增加趋势显著,达20%～30%。     

导线覆冰极值的概率分布模拟及其应用试验

利用南方地区多个气象站和电力部门观冰站的导线覆冰逐日冰厚资料,将广义极值分布和广义帕雷托分布引入导线覆冰的概率模型研究中,通过超门限覆冰次数的泊松分布拟合检验,结合H ill图解,提出了基于超门限峰值法门限值的确定方法;对两种分布在导线覆冰极值模型拟合的适用性研究表明,广义帕雷托分布对各站覆冰冰厚极值的拟合精度最高;重现期冰厚极值估计随样本长度的变化分析表明,广义帕雷托分布模型极值估计的稳定性比广义极值分布强,一般样本容量达到25 a左右时,广义帕雷托分布重现期冰厚极值的估计趋于稳定,可以作为短序列下估计导线覆冰极值的较好方法。     

中国南方夏半年湿期概率特征及其极值风险分析

以中国南方诸代表站近40 a(1965—2004年)夏季(5—9月)逐日降水资料为研究对象,探讨了夏半年各站湿期游程及其极端值的概率分布最佳模式。在对各站湿期游程分别验证指数分布的基础上,作耿贝尔(Gumbel)极值分布和广义帕雷托分布(GPD)拟合,进而对两者的拟合效果进行比较。并由此对湿期长度估计其不同重现期(如20 a一遇、50 a一遇和100 a一遇)的极端湿期长度的分位数概率。经K-S方法的统计检验,证明GPD分布拟合效果较好,能更加精确的模拟出中国南方夏季的极端连雨日数及其概率。     

1961-2014年新疆降水极值概率特征及拟合不确定性分析

本文采用多种函数初步探讨了1961—2014 年新疆地区日降水极值概率拟合分析中的不确定性。结果表明：新疆气候在暖干向暖湿转变背景下,更易出现极端强降水事件,由此影响极端降水拟合的不确定性;在统计函数选择以及重现期极值拟合中存在不确定性,因此建议在进行新疆区域降水极值分析时选择整体表现较好的GEV、Gen.Logistic、Log-Pearson 3、Pearson 6 和Wakeby 函数进行合成分析,在进行单站分析时给出具有界限范围的降水极值综合曲线图,以此减少来自函数方面的不确定性。     

A newly-discovered GPD-GEV relationship together with comparing their models of extreme precipitation in summer

It has been theoretically proven that at a high threshold an approximate expression for a quantile of GEV （Generalized Extreme Values） distribution can be derived from GPD （Generalized Pareto Distribution）. Afterwards, a quantile of extreme rainfall events in a certain return period is found using L-moment estimation and extreme rainfall events simulated by GPD and GEV, with all aspects of their results compared. Numerical simulations show that POT （Peaks Over Threshold）-based GPD is advantageous in its simple operation and subjected to practically no effect of the sample size of the primitive series, producing steady high-precision fittings in the whole field of values （including the high-end heavy tailed）. In comparison, BM （Block Maximum）-based GEV is limited, to some extent, to the probability and quantile simulation, thereby showing that GPD is an extension of GEV, the former being of greater utility and higher significance to climate research compared to the latter.     

广义帕雷托分布在重庆暴雨强降水研究中的应用

引进广义帕雷托分布(GPD),借助于现代L-矩估计方法,模拟重庆地区极端降水事件,推算一定重现期的极端降水量分位数。模拟试验表明,基于超门限峰值法(POT)的GPD不但计算简便,而且基本不受原始序列样本量的影响,具有全部取值域的高精度稳定拟合(包括高端厚尾部),与GEV模拟结果相比,GPD具有更高精度和稳定性,更为实用。     

Regional analysis of annual maximum rainfall using TL-moments method

Information related to distributions of rainfall amounts are of great importance for designs of water-related structures. One of the concerns of hydrologists and engineers is the probability distribution for modeling of regional data. In this study, a novel approach to regional frequency analysis using L-moments is revisited. Subsequently, an alternative regional frequency analysis using the TL-moments method is employed. The results from both methods were then compared. The analysis was based on daily annual maximum rainfall data from 40 stations in Selangor Malaysia. TL-moments for the generalized extreme value (GEV) and generalized logistic (GLO) distributions were derived and used to develop the regional frequency analysis procedure. TL-moment ratio diagram and Z-test were employed in determining the best-fit distribution. Comparison between the two approaches showed that the L-moments and TL-moments produced equivalent results. GLO and GEV distributions were identified as the most suitable distributions for representing the statistical properties of extreme rainfall in Selangor. Monte Carlo simulation was used for performance evaluation, and it showed that the method of TL-moments was more efficient for lower quantile estimation compared with the L-moments.     

Extreme value analysis in the Danube lower basin discharge time series in the twentieth century

The daily discharge time series in the lower Danube basin (Orsova) have been considered for the 1900–2005 period. The extreme value theory (EVT) is applied for the study of daily discharges incorporating some covariates. Two methods are applied for fitting the data to an extreme value distribution: block maxima and peaks over thresholds (POT). Using the block maxima approach associated with the use of the generalised extreme value (GEV) distribution, monthly and seasonal maxima of daily discharge for 1900–2005 have been analysed. Separately the monthly maxima of daily discharge for the 1958–2001 was analysed in order to be compatible with atmospheric circulation available from ERA-40. For performing parameter estimation, the maximum likelihood estimation (MLE) method was used. From the three possible types of GEV distribution, a Weibull distribution fits both the monthly and seasonal maxima of the daily discharges very well. The North Atlantic Oscillation (NAO) and the first ten principal components (PC) of the decomposition in multi-variate empirical orthogonal functions (MEOF) of three atmospheric fields (sea level pressure, 500 hPa and 500–1000 hPa thickness) over the Atlantic-European region (ERA-40), have been introduced as covariates. An improvement over the model without the covariate is found by incorporating NAO as the covariate in location parameter, especially for the spring maxima having the NAO as predictor during the winter. Related to atmospheric circulation influence, the most significant results are obtained by incorporating the first 10 PCs of the MEOF in the location parameter of GEV distribution within a month before the month of the discharge level. Regarding the POT approach associated with generalised Pareto distribution (GPD), different thresholds have been tested for daily discharges in the period 1900–2005, where the maxima were fitted by a bounded (or beta) distribution.     

Using the quantile regression method to analyze changes in climate characteristics

Possibilities to use the non-parametric regression analysis method, named the quantile regression, for the estimation of changes in climate characteristics are considered. When analyzing the trends of climatic series, the quantile regression method enables to get the information on trends along the whole range of quantile values from 0 to 1 of dependent variable distributions, that is more informative than the use of traditional regression technique, based on the least-squares method (LSM) and enabling to obtain trend estimations for average values of the dependent variable only. Trend estimation errors for various methods are analyzed. The computation of quantile regression parameters for real climatic series is executed. Series of meteorological variables of the diurnal resolution, which characterize the surface climate (minimal, average, and maximal diurnal temperatures) and free atmosphere climate (temperature of isobaric surfaces up to 30 hPa inclusive) are considered. Seasonal peculiarities in trend manifestation at different parts of quantile range of these meteorological values are discussed. Concerning the problem of the analysis of climate trends, the quantile regression method seems to be perspective from the point of view of more detailed understanding of processes in the climate system, such as the surface and tropospheric warming, stratospheric cooling, long-period changes in characteristics of climate variability and extremity.     

Statistical analysis of annual maximum rainfall in North-East India: an application of LH-moments

An attempt has been made to determine the best fitting distribution to describe the annual series of maximum daily rainfall data for the period 1966 to 2007 of nine distantly located stations in North East India. The LH-moments of order zero (L) to order four (L₄) are used to estimate the parameters of three extreme value distributions viz. generalized extreme value distribution (GEV), generalized logistic distribution (GLD), and generalized Pareto distribution (GPD). The performances of the distributions are assessed by evaluating the relative bias (RBIAS) and relative root mean square error (RRMSE) of quantile estimates through Monte Carlo simulations. Then, the boxplot is used to show the location of the median and the associated dispersion of the data. Finally, it can be revealed from the results of boxplots that zero level of LH-moments of the generalized Pareto distribution would be appropriate to the majority of the stations for describing the annual maximum rainfall series in North East India.     

未来情景下南水北调中线工程水源区极端降水分布特征

利用南水北调中线工程水源区9个气象站点1961-2008年的日降水资料和IPCC第四次评估报告多模式数据结果,抽取逐年的最大日降水量序列样本,运用广义极值分布（GEV）和广义帕累托分布 (GPD）两种极值统计模型对样本进行拟合,遴选出描述流域最大日降水量分布规律的最优概率模型,推算重现期对应的降水量值,并预估该流域极端降水事件在未来气候变化情景下的响应。研究表明：南水北调中线工程水源区降水极值均符合GEV和GPD分布,但GPD模型更适合用于描述该流域降水极值分布;未来气候变化情景下用GPD分布拟合的降水极值优于使用GEV分布;A2情景下极端降水事件的发生将更频繁、更强烈,A1B情景下次之,B1情景下相对较小,表明未来高排放气候情景对极端降水事件的影响比中、低排放情景大。     

The High Concentration Tails Of The Probability Density Function Of A Dispersing Scalar In The Atmosphere

The distribution function for concentrations of a scalar pollutant dispersing in the turbulent atmosphere has a finite domain that is bounded above and below. Three methods, based on extreme value statistics, are used to obtainestimates for the upper bound and to describe the high concentration tailbehaviour of the distribution; all three methods are applied to concentrationdata obtained from experimental atmospheric releases. Quantile quantile (QQ)plots are used to assess the goodness of fit of the resulting estimates of thedistribution, and also to compare the performance of the three methods. Thepredicted values for the upper bound are orders of magnitude less than thesource concentration, illustrating that molecular diffusion has a large effecton the high concentrations.