稳建回归的反复加权最小二乘迭代解法及其应用

用反复加权最小二乘迭代法求得稳健回归方程，将稳健回归方程与最小二乘的回归方程作了比较。结果表明，稳健使残差绝对值之和逐步减小并收敛。这种方法用于长江中下游降水预报，独立与非独立样本资料的检验表明，稳健回归方程比最小二乘回归效果好。     

Outlier Problem in Evaluating Similarity Functions in the Stable Atmospheric Boundary Layer

The gradient-based similarity approach removes turbulent fluxes as governing parameters and replaces them with vertical gradients of mean wind speed and potential temperature. As a result, the gradient Richardson number, Ri, appears as a stability parameter instead of the Monin–Obukhov stability parameter z/L (L is the Obukhov length). The gradient-based scaling is more appropriate for moderate and very stable conditions when the gradients are large and their errors are relatively small whereas z/L becomes ambiguous in these conditions because turbulent fluxes are small. However, the gradient-based formulation is faced with a problem related to the influence of Ri outliers: outliers with high values of Ri can exist in conditions that are really near-neutral. These outliers are mapped into the very stable range in plots in which Ri is the independent variable and may lead to spurious dependencies for bin-averaged data (spurious bin-averaging). This effect is quite large for functions that are steep for the gradient-based scaling. The present study uses the Surface Heat Budget of the Arctic Ocean (SHEBA) data to examine the problem and proposes two methods, conditional analysis and independent binning, to limit the influence of outliers on bin-averaging. A disadvantage of the conditional analysis is associated with eliminating outliers based on criteria that could be considered as subjective. The independent bin-averaging method does not have this disadvantage, but the scatter of the bin-averaged points is higher than for the conditional analysis, rendering data analysis and interpretation difficult.     

Integrative Monitoring Analysis Aiming at the Detection of Spatial and Temporal Trends of Metal Accumulation in Mosses

Goal and Scope. The UNECE heavy metals in Mosses Surveys provide data on the accumulation of metals in naturally growing mosses throughout Europe. Using Germany as an example, this article concentrates on the elaboration and application of methods of data analysis that are necessary for a comprehensive interpretation of spatial and temporal trends in metal accumulation.Methods and Results. The sampling of mosses, and the chemical analysis of Cr, Cu, Fe, Ni, Pb, Ti, V and Zn in 1990, 1995 and 2000 are summarized briefly. The variogram analyses detect distinct autocorrelation structures in the sample data which may consequently be estimated for all sites in Germany without monitoring metal accumulation by means of ordinary kriging. A procedure for the geostatistical detection of spatial outliers was developed and applied and, after the elimination of the spatial outliers, several measurements were seen to indicate an adequate quality of the geostatistical estimations. The cluster analyses of the z-transformed estimation data result in a nominal multi-element index that indicates regional metal accumulation types over time. Percentile statistics serve for computation of an ordinal, scaled, multi-element accumulation index which is spatially differentiated over time in terms of multivariate, statistically defined ecoregions. The integrative statistical analysis reveals, from 1990 to 2000, that the metal accumulation declines up to 80% in some of the ecoregions. Hot spots of metal accumulation are mapped and interpreted by means of metadata analysis.Conclusions. Dot maps depict the spatial structure of the metal accumulation without spatial bias. This information, detailed with respect to metal species and space, should be generalized for better supporting the interpretation. The combination of geostatistical analysis and estimation, percentile and multivariate statistics is suitable for the calculation of indices that serve for a comprehensive mapping of metal accumulation in the ecoregions over time, for quantifying the bias of the surface estimation, and for mapping spatial outliers and hot spots of metal accumulation.     

动态时变参数方法在云量精细化预报中的应用研究

利用2008年1月—2013年12月以及2017年1—11月全球天气预报系统（GFS）预报场资料,采用自适应线性最小二乘回归（LS）和自适应递推卡尔曼（Kalman）滤波两种动态时变参数方法,建立了河套周边地区0~168 h预报时效的总云量精细化预报,并与GFS模式直接输出的总云量、线性预报模型逐步回归预报方法得到的总云量以及非线性预报模型BP神经网络和最小二乘支持向量机回归方法（LSSVM）得到的总云量进行了对比,结果如下:（1）相比GFS模式直接输出的总云量,LS、BP神经网络、LSSVM得到的总云量与实况值的平均绝对误差均明显减小。LS方法误差最小,LS方法的年MAE均在20%~25%,且随着预报时效的延长,改进效果越大。LS方法、多元逐步回归方法、BP神经网络、LSSVM四种方法在6—8月的改进效果最大。（2）LS方法预报的总云量与实况云量的相关性最好,即使168 h预报时效的相关系数依然在0.64以上,远高于其他几种模型的预报结果。（3）LS方法能够明显地提高少云和多云天空状况下预报的击中率,且最优（少云击中率平均提高24 %,多云击中率平均提高34 %）。（4）自适应递推Kalman滤波方法存在预报滞后现象,改进效果不明显。      

A machine learning approach to quality-control Argo temperature data

A machine learning approach is proposed to identify temperature outliers from Argo float profiles as a complementary procedure to current Argo quality control. A machine learning unsupervised classification (i.e., the Gaussian mixture model, GMM) is applied to cluster the Argo data into classes to construct convex hulls with the smallest polygons encompassing all the data points. Good or bad temperature data are identified as within or outside the polygons based on point-in-polygon analysis implemented by the ray casting algorithm. The South China Sea was selected as an example and results showed that the proposed approach could identify more than 70% of the profiles containing the outliers and mark the outliers automatically at the same time. This highlights the potential of the proposed methodology to be a good complementary quality control method.摘要本文提出了一种基于机器学习的Argo浮标温度异常值检测方法. 该方法采用机器学习无监督算法高斯混合模型对Argo浮标数据进行聚类分析, 并构建包围所有数据点的最小多边形的凸包. 基于射线投影算法实现点在多边形内分析, 通过自动识别数据点位于凸包内外来判断该数据点数据质量的好坏. 本文采用南海区域Argo浮标数据对该方法进行测试, 结果表明该方法可以识别70%以上的包含异常值的温度剖面, 同时自动标记出各异常值点.     

FY-3A气象卫星臭氧总量数据的质量控制方案及其在台风Tembin(2012)和Isaac(2012)中的应用

我国第二代极轨气象卫星“风云三号”A星(FY-3A)上搭载的紫外臭氧总量探测仪(Total Ozone Unit,TOU)每天可以提供一次覆盖全球的臭氧总量观测。为了在数值预报中应用TOU的臭氧资料,从资料同化角度发展了一套质量控制方案。首先基于臭氧总量和平均位势涡度的高相关性建立了逐日动态更新的臭氧线性回归预报模型,然后使用双权重算法对臭氧资料进行质量控制。将该质量控制方案应用于台风Tembin(2012)和Isaac(2012)个例,试验结果说明该方案可以体现出臭氧总量和平均位势涡度之间相关关系的逐日变化,识别出的离群资料百分比随时间变化较稳定,可以保留原始资料的主体信息,并且显著降低了原始资料的标准差。同时,质量控制后的臭氧数据与统计拟合量更加一致,观测减拟合的概率密度函数分布形式也更接近高斯分布,有利于后续的资料同化。     

GRAPES全球三维变分同化中卫星微波温度计亮温的背景误差及在质量控制中的应用

模式变量背景误差在观测空间的投影,也即观测变量的背景误差包含了变分同化系统的重要信息,其在诊断和分析变分同化系统中资料的影响等方面具有重要作用,特别是在背景场检查质量控制中。在GRAPES全球三维变分同化(3DVar)系统中仅给定了控制变量的背景误差,并未直接给定观测变量的背景误差。为了能够对GRAPES全球3DVar进行全面的诊断和分析,改进卫星微波温度计资料的质量控制,推导出GRAPES全球3DVar同化系统控制变量随机扰动方法估计观测变量的背景误差的公式,为分析和改进GRAPES全球3DVar提供了一个有力工具,并进而估计了AMSU-A亮温的背景误差,分析了AMSU-A不同通道亮温的背景误差特征,将其应用于GRAPES全球3DVar的AMSU-A亮温的背景场检查质量控制中。结果表明,控制变量随机扰动方法估计的GRAPES全球3DVar同化系统AMSU-A亮温的背景误差正确合理。同化循环预报试验结果表明,亮温的背景误差在背景场检查中的应用显著提高了GRAPES全球3DVar同化的亮温资料的数量,显著提高了GRAPES南半球对流层中高层位势高度场的预报技巧。在GRAPES全球3DVar同化系统中推导和实现的控制变量扰动方法为诊断和分析GRAPES全球3DVar观测资料同化效果提供了有力工具。      

Impact of different regression frameworks on the estimation of the scaling properties of radar rainfall

Rainfall is characterized by high variability both in space and time. Despite continuous technological progress, the available instruments that are used to measure rainfall across several spatio-temporal scales remain inaccurate. To remedy this situation, scaling relationships of spatial rainfall offer the potential to link the observed or predicted precipitation quantities at one scale to those of interest at other scales. This paper focuses on the estimation of the spatial rainfall scaling functions. Standard scaling analysis constructed by means of the ordinary least squares method often violates such basic assumptions implicit in its use and interpretation as homoschedasticity, independence, and normality of the errors. Consequently, the authors consider alternative regression frameworks i.e. bootstrapping regression, semi parametric linear model, and multilevel normal linear model to show how these different approaches exert a significant impact on the multifractal analysis of radar rainfall. In addition, the uncertainties associated with the construction of the scaling function due solely to the regression procedure are quantified. The radar data come from the polarimetric C-band weather radar located in Rome, Italy, and the scaling properties are computed for a square domain centred on the radar site with a side length of 128 km and a finest resolution of 1 km².     

Anticipated changes in analysis accuracy with the removal of Ocean Weather Ship P

Abstract

Analysis error patterns have been established for the Pacific Weather Centre Experiment Area, and comparisons made between errors computed for meteorological observing arrays, including Ocean Weather Ship (OWS) P, and errors computed for several alternative arrays which excluded OWS P. These assessments of the impact of replacing the ocean weather ship with alternative observing equipment indicate that, above the 1000‐mb pressure surface, there will be a significant loss of accuracy in the forecast‐minus‐observation analyses regardless of proposed additional report systems. Near the surface, forecast error variances are estimated to decrease slightly with an increase of reports from buoys and ships of opportunity within the region.

The dependence of the assessments on the data selection procedure and on correlation representations for the region suggest that some loss may be compensated by more efficient use of available data. Refinements in the objective analysis scheme are seen to be especially important to analysis accuracy in regions lacking radiosonde coverage.     

离散型应变量的回归诊断和离散回归法

本文从大量天气预报实例的普查结果,归纳了离散型应变量的回归残差分布模式。回归诊断表明模式的单向性是由一些高杠杆点引起。对这些高杠杆点的进一步分析可以看到,它们的残差存在着统计预报意义上的不合理性,不合理性又造成回归系数LS估计的误差。针对这些问题,本文提出了离散回归法。     

Data-Assimilative Hindcast and Climate Forecast of Storm Surges with an ASGF Regression Model

ABSTRACT

This study demonstrates that long-term climate model solutions can be efficiently converted to storm surge time series at points of interest (POIs) for the future. The all-source Green's function (ASGF) regression model is used for this conversion. In addition to being data assimilative, the ASGF regression model can also simulate storm surges at a POI faster than the traditional modelling approach by orders of magnitude. This is demonstrated using the tidal gauge at Sept-Îles (Quebec, Canada) in the Gulf of St. Lawrence as the POI. First the ASGF regression model is used to assimilate 32 years of tidal gauge data, producing a continuous hindcast of storm surges and a set of best-estimate regression parameters. Second, the ASGF regression model with the best-estimate parameters is used to convert a Canadian Regional Climate Model solution (CRCM/AHJ) to an hourly time series of storm surges from 1961 to 2100. Gumbel's extreme value analysis (EVA) is then applied to the time series as a whole and also to tri-decadal segments. The tri-decadal approach is used to investigate whether there is any progressive shortening or lengthening of storm surge return periods as a result of future climate change. A method for correcting for bias due to the forcing field at the EVA level is also demonstrated.     

The Impact of Assimilating FY-3C GNOS GPS Radio Occultation Observations on GRAPES Forecasts

In the present study, a gross quality control (QC) procedure is proposed for the Global Navigation Satellite System Occultation Sounder (GNOS) Global Positioning System radio occultation (GPS RO) refractivity data to remove abnormal data before they are assimilated. It consists of a climate extreme check removing data outside the range of the Constellation Observing System for Meteorology, Ionosphere and Climate (COSMIC) climate maxima and minima over approximately five years, and a vertical gradient check that rejects profiles containing super-refractions. These two QC steps were applied sequentially to identify outliers in GNOS GPS RO refractivity data during boreal winter 2013/2014. All of the abnormal refractivity profiles and the outliers at each level of the GNOS GPS RO observations were effectively removed by the proposed QC procedure. The post-QC GNOS GPS RO refractivity observations were then assimilated in the Global/ Regional Analysis and PrEdiction System (GRAPES) using the three-dimensional variational (3D-Var) system. The impacts of the GNOS refractivity observation on GRAPES analysis and forecasting were evaluated and analyzed using an observation system experiment run over one whole winter season of 2013 / 2014. The experiment results demonstrated a positive impact of GNOS GPS RO data on analysis and forecast quality. The root mean squared error of GRAPES analysis temperature was reduced by 1% in the Southern Hemisphere (SH) extratropics and in the tropics, and the anomaly correlation scores of the forecasted 500-hPa geopotential height over the SH increased significantly during days 1 to 5. Overall, the benefits of using GNOS GPS RO data are significant in the SH and tropics.     

评估FY-3A微波湿度计O-B对云的识别能力

应用FY-3A微波湿度计2010年1月份的Level-1c观测亮度温度O,NCEP GFS 6 h的预报场作为背景场,用RTTOV 9.3版本辐射传输模式模拟的亮度温度B以及美国NOAA-18 MHS业务微波地表和降水产品,研究了双权重质量控制算法对FY-3A MWHS 通道3至5云和降水视场的识别能力。研究表明双权重质量控制算法判断的负观测增量O-B的离群点中,大多数都受云和降水影响。通道3约占60%,通道4约80%,通道5超过80%。当降水率大于0.2 mm·h^-1时,通道3负离群点可识别超过60%降水云,随着降水率增大识别率超过80%。而通道4对大于0.2 mm·h^-1的降水的识别率超过90%。通道5负离群点几乎可以剔除100%的降水影响资料。在目前还没有MWHS自身云检测产品的条件下,双权重质量控制算法可剔除大部分云和降水影响视野。     

A Note on an improved expression for the atmospheric CO2, N2O,CH4 and O2 integral transmission function: Research note

Abstract

A new analytical expression for the integral transmission function corresponding to the uniformly mixed atmospheric gases (CO₂, N₂O, CH4 and O₂) is proposed for solar radiation models. The expression is based on the latest known spectral absorption data of the gases as well as on the Ñeckel and Labs (1981) extraterrestrial solar spectrum, which is the best currently available. The proposed formula is compared with the known expressions for the integral transmission function of the uniformly mixed gases and the results of this comparison are discussed.     

Frequency analysis of short‐duration rainfalls

Abstract

An investigation was carried out to determine the appropriateness of a probability distribution and fitting technique commonly used in Canada for rainfall frequency analysis.

The extreme value type 1 (EV1) distribution was assessed for three long‐term Canadian stations: Victoria, St Thomas and Québec. The EV1 distribution appears to provide a reasonable fit for durations varying from 5 min to 6 h, but is not clearly superior to another two‐parameter distribution, the lognormal.

The fitting technique, known as modified moments or regression, was assessed by comparing it with three other fitting techniques: moments, maximum likelihood and adjusted maximum likelihood. This comparison was carried out using Monte Carlo simulation techniques over the parameter space deemed to be representative of short duration rainfall data for Canada. In terms of estimating rainfall amounts of specified return period, the modified moments technique was the poorest with regard to both bias and efficiency. In general, the maximum likelihood estimates were the most efficient and were relatively unbiased.     

Theoretical meteorology and the operational meteorologist

Abstract

Values of incoming solar and long‐wave radiation measured at the vessel Quadra during the three phases of GATE are used to assess the daily performance of three models, one for solar and two for long‐wave radiation. The solar radiation model, which uses data on precipitable water and cloud amount at three levels in the atmosphere performed satisfactorily during the first phase but gave poor results in the other two phases when cumulonimbus became more dominant. Both the flux‐emissivity approach using measured and interpolated Upper air data and Paltridge's empirical procedure produced estimates of long‐wave radiation which compared very closely with the measurements.     

Solar energy system design: The impact of mesoscale variations in solar radiation

Abstract

A 12‐station network located in and around Vancouver, B.C. has been used to define the local (mesoscale) variations in solar energy. Numerical models are used to convert the horizontal data to inclined surface irradiances which, along with temperature, are the inputs required by the simulation models used to assess the significance of these variations on the performance and economics of both active and passive solar heating systems.

Solar radiation differences over the study area have a marked influence on the operation of a domestic hot water heating system while the temperature variations between the same stations have a comparatively small impact. For both active and passive space heating the situation is reversed because temperature has a large impact on the energy demand and hence on the extent of solar energy utilization. The results suggest that local variations of solar radiation should be incorporated in solar domestic hot water design studies while spatial variations in temperature should receive relatively more attention in space heating studies.     

The local spline vertical interpolation method of temperature and geopotential height fields and the time-dependent difference form of the hydrostatic equation

In numerical weather prediction (NWP), the accuracy of vertical interpolation of the initial data is a problem which is greatly concerned by people. In this paper, we specify vertical distributions of the temperature and the geopotential height fields and examine three interpolation methods, i.e. the Lagrangian polynomial inter-polation method (hereafter abbreviated to LP method), the linear interpolation method (LN method) and the local spline interpolation method (LS method) proposed by the author. The examination shows that when the vertical resolution of the initial data is high enough, for example, the number of the given data levels N is 10 or more, all the three methods get good accuracy of interpolation, especially, the LP and the LS methods have very little errors almost tending to zero, while the LN method has a little larger errors than the two formers and the errors at various levels have the same sign. When N is reduced to 5, the LP and the LS methods still have quite good accuracy and similar error distributions, while the LN method has less accuracy. If the geopo-tential height field needs to be adjusted in order to satisfy the hydrostatic equilibrium with the temperature field which is assumed fixed, then the LS method has minimum errors. The examination also indicates that the vertical resolution with at least 5 levels of initial data can keep the interpolation accuracy. Otherwise the accuracy will not be guaranteed no matter which method is used.It is also pointed out in this paper that the temperature and the geopotential height fields can be given inde-pendently in numerical prediction models in order to keep higher interpolation accuracy. However, the hydro-static equation should be finite differenced in other way which is somewhat different from the conventional one. In other words, the time dependent difference form of the equation should be used, so that the initial interpola-tion accuracy could have influence on the time integration.     

Estimation of air temperature at 5 cm above grassland at the Zagreb-Maksimir Observatory

Summary The possibility of estimating air temperature at 5 cm above grassland is examined using an empirical, as well as a theoretical approach. For this purpose, hourly data (air temperature, cloudiness, air pressure and wind speed) at the Zagreb-Maksimir Observatory (Croatia) for the period June–August 1975 have been used. A clustering procedure has been applied to form groups of days having similar diurnal variations in air temperature difference between 5 cm and 2 m. This procedure has identified, three groups of similar days (types). The air temperature at 5 cm above grassland has been estimated using multiple regression and a method which is based on the Monin–Obukhov similarity theory. The results obtained are more accurate for near-neutral stability conditions, i.e. during cloudy weather (Type 1), and the least accurate in the case of a clear sky (Type 3). Through clustering, systematic errors have been discovered in these approaches, which are briefly discussed.     

辅助模型辨识方法(5):最小二乘辨识

借助于辅助模型辨识思想,针对白噪声干扰的输入非线性有限脉冲响应系统,研究了辅助模型最小二乘辨识方法、辅助模型多新息最小二乘辨识方法、变递推间隔辅助模型最小二乘辨识方法、变递推间隔辅助模型多新息最小二乘辨识方法、等递推间隔辅助模型多新息最小二乘辨识方法,以及有限数据窗最小二乘辨识方法,包括引入加权因子（加权矩阵）、遗忘因子得到的一些相应辨识方法.