基于机器学习的数值天气预报风速订正研究

对风速进行准确预测是精细化天气预报服务(如风能发电、冬季奥运会赛场条件保障等)的重要环节。本文基于三种机器学习算法(LASSO回归、随机森林和深度学习),对数值天气预报模式ECMWF预测的华北地区近地面10 m风速进行订正。首先利用LASSO回归算法提取对10 m风速有重要影响的气象要素特征集,将其作为三种机器学习算法的输入,建立相应模型对ECMWF预测的风速进行订正。用提取后的气象要素特征集建模有助于减少计算量和存储开销,并减小模型的复杂性,从而提高模型的泛化能力。将订正结果与传统订正方法模式输出统计(model output statistics,MOS)得到的订正结果进行对比。结果表明,三种机器学习算法的订正效果均好于MOS方法,显示了机器学习方法在改善局地精准气象预报方面的潜力。     

京津冀站点风温湿要素的机器学习订正方法

基于线性回归方法、梯度提升回归方法(GBRT方法)、XGBoost方法和堆叠集成学习方法(Stacking方法)4种机器学习方法,采用误差分析建模思路,针对北京城市气象研究院研发的睿图-睿思系统对2020年12月—2021年11月所有起报时次未来3~12 h的2 m温度、2 m相对湿度、10 m风速以及10 m风向4种气象要素预报,开展京津冀复杂地形下的站点预报误差订正技术研究及试验应用。结果表明:基于预报误差分析构建的4种订正模型中,由于Stacking方法集成了前3种方法的优势,在4个季节的4种气象要素订正中均表现最佳,其他3种单一机器学习方法试验中,XGBoost方法表现最佳,其后依次为GBRT方法、线性回归方法,但均对预报准确率有明显的正向提升效果。总体上,基于机器学习方法构建的预报误差订正模型可有效降低系统原始预报误差,有助于进一步提升复杂地形下站点客观释用产品的预报准确性。     

一个简单的格点温度预报订正方法

格点气象要素预报是中国气象局目前的主推业务和未来天气预报的发展方向。本文基于欧洲中期数值预报中心ECMWF高分辨率模式2 m温度预报资料,在传统中央气象台站点温度指导预报SCMOC和回归方法建立的站点温度预报的基础上,提出"站点订正值向格点传递"的方法来订正格点温度预报。结果表明:(1)SCMOC站点最高、最低温度24~168 h预报误差2℃准确率分别平均高于ECMWF的10.0%和23.1%,ECMWF存在较大的系统性偏差,最低温度预报偏高,最高温度预报偏低。(2)"站点订正值向格点传递"方法能够订正模式格点温度预报的系统误差,且整体上不改变原ECMWF温度预报场的空间形态和原模式对地形的刻画特征。(3)利用研究区域内98个县级站SCMOC温度预报,订正ECMWF格点场,返回到区域内1289个乡镇站进行检验,结果24 h最低、最高温度1℃的准确率较ECMWF分别提高22.8%和11.9%,2℃的准确率则分别提高29.7%和17.4%。最低(高)温度绝对误差平均减小0.99℃(0.69℃),平均误差(ME)下降到0.7℃(-0.9℃)以内。(4)通过一元线性回归,得到98个县级站的温度预报,返回差值场来订正格点场,也能较好地订正ECMWF的系统性误差。对比两种方法,SCMOC差值传递在最低温度订正方面有较大的优势,而回归方法的最高温度订正效果较好。此外,回归方法能够较好地改善逐时温度预报效果。该方法已成功运用于陕西省精细化格点预报业务系统中。     

基于机器学习的模式温度预报订正方法研究

机器学习在气象数值模式的后处理中表现优越,但其稳定性和适用性有待深入探究。本文选取了ECWMF模式包括2米温度、风、降水等多气象要素预报产品和安徽省80个国家气象站观测2米温度实况资料,分析了EC模式在安徽省站点温度预报误差,利用决策树、随机森林、LightGBM三种机器学习算法订正EC 模式0-72小时温度站点预报,并将其与传统MOS订正方法和SPCC主观预报产品进行了对比。结果表明：EC模式高温预报误差明显高于低温预报,在安徽皖南山区和大别山区存在较大误差;机器学习算法中最高温度预报随机森林表现最优,最低温度预报LightGBM最优,比EC模式平均绝对误差MAE分别降低了0.55℃、0.2℃,均方根误差RMSE分别降低0.6℃、0.31℃,预报准确率提高了18.16%和5.19%;高山站独立建模并融合周围站的信息能有效降低模型误差;相比SPCC主观预报产品,机器学习预报模型在高温和寒潮过程中互有优劣,但在天气转折初期落后;机器学习可以作为常规预报模式的补充,能显著优化或改善传统预报中温度预测精度,特别是对于数据缺乏的高山站点。     

镇江地区ECMWF模式高温预报订正方法研究

针对镇江ECMWF模式168 h内高温(t≥35℃)预报结果提出四种后处理订正方案,包括一元线性回归法、差值法、综合法和递减平均法;借助均方根误差等四种检验方法就订正效果进行评估,找寻最优订正方案。结果表明,四种订正方法都明显改善了ECMWF模式高温预报,订正后的均方根误差、平均绝对误差及最大绝对误差较订正前均有所减小,预报准确率显著提高。对于24 h时效内预报,四种订正方法各有优势。对于48～168 h时效预报,一元线性回归法效果更优。采用分时效对ECMWF模式高温预报结果进行后处理,考虑24 h预报订正使用递减平均方法,48～168 h预报订正使用一元线性回归法,可以更大程度地提高预报准确率。     

基于SVD与机器学习的华南降水预报订正方法

降水是在多种天气系统和复杂物理过程共同影响下形成的,因此降水预报难度较大。由于数值预报模式的局限性,使得模式预报产品存在一定误差。为探讨更加有效的模式预报产品误差订正方法,基于奇异值分解（SVD）与机器学习（多元线性回归、套索回归、岭回归）构建订正模型,对2007—2019年4月1日—6月30日华南前汛期欧洲中期天气预报中心（EC）模式降水预报产品进行误差订正试验。结果表明:基于SVD与机器学习相结合的订正模型能有效降低EC模式降水预报产品在华南的预报误差,均方根误差最大优化率达4.2%,累计超过69%的站点得到不同程度的优化;SVD与机器学习相结合的订正模型能很好地处理因子间共线性问题,具有更好的鲁棒性;而对多个订正模型加权集成,均方根误差优化率达5.7%,累计超过77%的站点得到优化,显然加权集成方法订正效果不仅优于EC模式预报产品,也优于参与集成的任一订正模型。     

基于ECMWF产品的站点气温预报集成学习误差订正

提出一种基于数值模式预报产品的气温预报集成学习误差订正方法,通过人工神经网络、长短期记忆网络和线性回归模型组合出新的集成学习模型（ALS模型）,采用2013—2017年的欧洲中期天气预报中心数值天气预报模式2 m气温预报产品和中国部分气象站点数据,利用气象站点气温、风速、气压、相对湿度4个观测要素,挖掘观测数据的时序特征并结合模式2 m气温预报结果训练机器学习模型,对2018年模式2 m气温6~168 h格点预报产品插值到站点后的预报结果进行偏差订正。结果表明:ALS模型可将站点气温预报整体均方根误差由3.11℃降至2.50℃,降幅达0.61℃（19.6%）,而传统的线性回归模型降幅为0.23℃（8.4%）。ALS模型对站点气温预报误差较大的区域和气温峰值预报的订正效果尤为显著,因此,集成学习方法在数值模式预报结果订正中具有较大的应用潜力。     

基于机器学习的风场预报订正方法研究

为了实现更准确的站点风预报,结合中尺度数值模式WRF预报结果和自动气象站观测数据,采用反距离加权插值法,将模式网格和观测站点的数据进行融合构建训练集,利用3种机器学习方法对WRF预报的风场结果进行订正,优化风场预报准确率。其中随机森林模型实现风速的预报均方根误差(RMSE)平均降低了18.22%,风向降低了15.97%;LightGBM模型对于风速、风向的RMSE平均降低了18.60%和17.56%;深度神经网络模型对于风速、风向的RMSE平均降低了5.53%和9.10%。对2020年宁波市9个大风过程进行检验,利用LightGBM模型对于3个站点预报进行订正,结果表明风速的RMSE从4.61 m/s下降到2.14 m/s,平均降低了53.58%,风向的RMSE从30.31°下降到18.20°,平均降低了39.95%。     

基于CNN和RF算法的ECMWF降水分级订正预报方法

利用卷积神经网络(CNN)和随机森林回归模型,提出了一种新的欧洲中期天气预报中心(ECMWF)降水订正预报方法。该方法首先根据ECMWF模式对站点雨量预报值所属的等级进行划分,再计算出不同等级相对应的高相关因子矩阵。进一步利用CNN模型对高相关矩阵进行综合特征提取的学习和训练。最后对CNN模型最终输出的特征因子中,选取若干个与预报站点相关性高的特征,并与ECMWF降水量场插值到预报站点的因子一起,作为随机森林回归模型的输入因子进行预报建模。通过对10个预报试验站点未来24h降水量的分级和不分级订正预报试验,结果表明:(1)ECMWF降水量分级订正预报方法的平均绝对偏差和均方根误差分别比利用ECMWF插值到站点的预报方法减小了20%和15%;(2)24h暴雨及以上的降水分级订正预报方法的平均TS评分为0.32,也显著高于EC插值的0.19;(3)与利用同样的预报模型对全样本(不分级)的传统数值预报模式产品订正预报方法相比,本文提出的分级订正预报方法在总体预报精度和暴雨及以上的强降水预报TS评分上均有更高的预报技巧。     

基于Parsivel天气现象仪资料的郑州“7·20”罕见特大暴雨微物理特征分析

基于2016-2018年ECMWF模式温度预报和浙江省72个国家基本站观测资料,根据温度日变化特征,采用K-近邻(KNN)回归算法进行误差订正,改进浙江省172 h精细化温度预报。在KNN回归算法中,将模式起报时刻的温度视作“背景”,由模式预报减去起报时刻温度消除“背景”影响,得到温度日变化曲线,通过温度日变化曲线构建差异指标,选取历史相似个例。根据历史相似个例的误差特征,对温度预报进行订正,得到改进的温度预报。检验结果表明,KNN方案的温度预报平均绝对误差较ECMWF和30 d滑动平均误差订正方案(OCF)的分别减小26.2%和5.2%;日最高和最低温度预报误差绝对值小于2℃,准确率较ECMWF的分别提高14.8%和4.3%,较OCF的分别提高3.0%和1.3%。KNN方案对地形复杂地区的温度预报改进效果更为明显,对冷空气活动和夏季高温等天气过程预报改善效果也较稳定。     

COMPARATIVE ANALYSIS OF VARIOUS DATA OF AIR–SEA TEMPERATURE DIFFERENCE AND ITS VARIATION ACROSS SOUTH CHINA SEA IN THE PAST 35 YEARS

Using the International Comprehensive Ocean-Atmosphere Data Set(ICOADS) and ERA-Interim data, spatial distributions of air-sea temperature difference(ASTD) in the South China Sea(SCS) for the past 35 years are compared,and variations of spatial and temporal distributions of ASTD in this region are addressed using empirical orthogonal function decomposition and wavelet analysis methods. The results indicate that both ICOADS and ERA-Interim data can reflect actual distribution characteristics of ASTD in the SCS, but values of ASTD from the ERA-Interim data are smaller than those of the ICOADS data in the same region. In addition, the ASTD characteristics from the ERA-Interim data are not obvious inshore. A seesaw-type, north-south distribution of ASTD is dominant in the SCS; i.e., a positive peak in the south is associated with a negative peak in the north in November, and a negative peak in the south is accompanied by a positive peak in the north during April and May. Interannual ASTD variations in summer or autumn are decreasing. There is a seesaw-type distribution of ASTD between Beibu Bay and most of the SCS in summer, and the center of large values is in the Nansha Islands area in autumn. The ASTD in the SCS has a strong quasi-3a oscillation period in all seasons, and a quasi-11 a period in winter and spring. The ASTD is positively correlated with the Nio3.4 index in summer and autumn but negatively correlated in spring and winter.     

CHARACTERISTICS AND TRENDS OF CLIMATIC EXTREMES IN CHINA DURING 1959–2014

The spatial and temporal variations of daily maximum temperature(Tmax), daily minimum temperature(Tmin), daily maximum precipitation(Pmax) and daily maximum wind speed(WSmax) were examined in China using Mann-Kendall test and linear regression method. The results indicated that for China as a whole, Tmax, Tmin and Pmax had significant increasing trends at rates of 0.15℃ per decade, 0.45℃ per decade and 0.58 mm per decade,respectively, while WSmax had decreased significantly at 1.18 m·s~(-1) per decade during 1959—2014. In all regions of China, Tmin increased and WSmax decreased significantly. Spatially, Tmax increased significantly at most of the stations in South China(SC), northwestern North China(NC), northeastern Northeast China(NEC), eastern Northwest China(NWC) and eastern Southwest China(SWC), and the increasing trends were significant in NC, SC, NWC and SWC on the regional average. Tmin increased significantly at most of the stations in China, with notable increase in NEC, northern and southeastern NC and northwestern and eastern NWC. Pmax showed no significant trend at most of the stations in China, and on the regional average it decreased significantly in NC but increased in SC, NWC and the mid-lower Yangtze River valley(YR). WSmax decreased significantly at the vast majority of stations in China, with remarkable decrease in northern NC, northern and central YR, central and southern SC and in parts of central NEC and western NWC. With global climate change and rapidly economic development, China has become more vulnerable to climatic extremes and meteorological disasters, so more strategies of mitigation and/or adaptation of climatic extremes,such as environmentally-friendly and low-cost energy production systems and the enhancement of engineering defense measures are necessary for government and social publics.     

IMPACT OF ATMOSPHERIC LOW-FREQUENCY OSCILLATION ON THE IMMIGRATION OF NILAPARVATA LUGENS(STL) IN HUNAN AND JIANGXI PROVINCES

Various features of the atmospheric environment affect the number of migratory insects, besides their initial population. However, little is known about the impact of atmospheric low-frequency oscillation(10 to 90 days) on insect migration. A case study was conducted to ascertain the influence of low-frequency atmospheric oscillation on the immigration of brown planthopper, Nilaparvata lugens(Stl), in Hunan and Jiangxi provinces. The results showed the following:(1) The number of immigrating N. lugens from April to June of 2007 through 2016 mainly exhibited a periodic oscillation of 10 to 20 days.(2) The 10-20 d low-frequency number of immigrating N. lugens was significantly correlated with a low-frequency wind field and a geopotential height field at 850 h Pa.(3) During the peak phase of immigration, southwest or south winds served as a driving force and carried N. lugens populations northward, and when in the back of the trough and the front of the ridge, the downward airflow created a favorable condition for N. lugens to land in the study area. In conclusion, the northward migration of N. lugens was influenced by a low-frequency atmospheric circulation based on the analysis of dynamics. This study was the first research connecting atmospheric low-frequency oscillation to insect migration.     

A COMPARATIVE ANALYSIS OF ATMOSPHERIC AND OCEANIC CONDITIONS BEFORE THE OCCURRENCE OF TWO TYPES OF EL NIO EVENTS

The atmospheric and oceanic conditions before the onset of EP El Ni?o and CP El Ni?o in nearly 30 years are compared and analyzed by using 850 hPa wind, 20℃ isotherm depth, sea surface temperature and the Wheeler and Hendon index. The results are as follows: In the western equatorial Pacific, the occurrence of the anomalously strong westerly winds of the EP El Ni?o is earlier than that of the CP El Ni?o. Its intensity is far stronger than that of the CP El Ni?o. Two months before the El Ni?o, the anomaly westerly winds of the EP El Ni?o have extended to the eastern Pacific region, while the westerly wind anomaly of the CP El Ni?o can only extend to the west of the dateline three months before the El Ni?o and later stay there. Unlike the EP El Ni?o, the CP El Ni?o is always associated with easterly wind anomaly in the eastern equatorial Pacific before its onset. The thermocline depth anomaly of the EP El Ni?o can significantly move eastward and deepen. In addition, we also find that the evolution of thermocline is ahead of the development of the sea surface temperature for the EP El Ni?o. The strong MJO activity of the EP El Ni?o in the western and central Pacific is earlier than that of the CP El Ni?o. Measured by the standard deviation of the zonal wind square, the intensity of MJO activity of the EP El Ni?o is significantly greater than that of the CP El Ni?o before the onset of El Ni?o.     

ANALYSIS OF “BIMODAL PATTERN” STORM ACTIVITY CHARACTERISTICS OF THE BAY OF BENGAL

Storms that occur at the Bay of Bengal (BoB) are of a bimodal pattern, which is different from that of the other sea areas. By using the NCEP, SST and JTWC data, the causes of the bimodal pattern storm activity of the BoB are diagnosed and analyzed in this paper. The result shows that the seasonal variation of general atmosphere circulation in East Asia has a regulating and controlling impact on the BoB storm activity, and the “bimodal period” of the storm activity corresponds exactly to the seasonal conversion period of atmospheric circulation. The minor wind speed of shear spring and autumn contributed to the storm, which was a crucial factor for the generation and occurrence of the “bimodal pattern” storm activity in the BoB. The analysis on sea surface temperature (SST) shows that the SSTs of all the year around in the BoB area meet the conditions required for the generation of tropical cyclones (TCs). However, the SSTs in the central area of the bay are higher than that of the surrounding areas in spring and autumn, which facilitates the occurrence of a “two-peak” storm activity pattern. The genesis potential index (GPI) quantifies and reflects the environmental conditions for the generation of the BoB storms. For GPI, the intense low-level vortex disturbance in the troposphere and high-humidity atmosphere are the sufficient conditions for storms, while large maximum wind velocity of the ground vortex radius and small vertical wind shear are the necessary conditions of storms.     

LOW-FREQUENCY CIRCULATION AND EXTENDED-RANGE FORECAST IN CONNECTION WITH AUTUMN EXTREME HEAVY RAINFALL PROCESS IN HAINAN

Observed daily precipitation data from the National Meteorological Observatory in Hainan province and daily data from the National Centers for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR) reanalysis-2 dataset from 1981 to 2014 are used to analyze the relationship between Hainan extreme heavy rainfall processes in autumn (referred to as EHRPs) and 10–30 d low-frequency circulation. Based on the key low-frequency signals and the NCEP Climate Forecast System Version 2 (CFSv2) model forecasting products, a dynamical-statistical method is established for the extended-range forecast of EHRPs. The results suggest that EHRPs have a close relationship with the 10–30 d low-frequency oscillation of 850 hPa zonal wind over Hainan Island and to its north, and that they basically occur during the trough phase of the low-frequency oscillation of zonal wind. The latitudinal propagation of the low-frequency wave train in the middle-high latitudes and the meridional propagation of the low-frequency wave train along the coast of East Asia contribute to the ‘north high (cold), south low (warm)’ pattern near Hainan Island, which results in the zonal wind over Hainan Island and to its north reaching its trough, consequently leading to EHRPs. Considering the link between low-frequency circulation and EHRPs, a low-frequency wave train index (LWTI) is defined and adopted to forecast EHRPs by using NCEP CFSv2 forecasting products. EHRPs are predicted to occur during peak phases of LWTI with value larger than 1 for three or more consecutive forecast days. Hindcast experiments for EHRPs in 2015–2016 indicate that EHRPs can be predicted 8–24 d in advance, with an average period of validity of 16.7 d.     

AN ATTRIBUTION ANALYSIS OF CHANGES IN POTENTIAL EVAPOTRANSPIRATION IN THE BEIJING–TIANJIN–HEBEI REGION UNDER CLIMATE CHANGE

Based on the measurements obtained at 64 national meteorological stations in the Beijing–Tianjin–Hebei (BTH) region between 1970 and 2013, the potential evapotranspiration (ET0) in this region was estimated using the Penman–Monteith equation and its sensitivity to maximum temperature (Tmax), minimum temperature (Tmin), wind speed (Vw), net radiation (Rn) and water vapor pressure (Pwv) was analyzed, respectively. The results are shown as follows. (1) The climatic elements in the BTH region underwent significant changes in the study period. Vw and Rn decreased significantly, whereas Tmin, Tmax and Pwv increased considerably. (2) In the BTH region, ET0 also exhibited a significant decreasing trend, and the sensitivity of ET0 to the climatic elements exhibited seasonal characteristics. Of all the climatic elements, ET0 was most sensitive to Pwv in the fall and winter and Rn in the spring and summer. On the annual scale, ET0 was most sensitive to Pwv, followed by Rn, Vw, Tmax and Tmin. In addition, the sensitivity coefficient of ET0 with respect to Pwv had a negative value for all the areas, indicating that increases in Pwv can prevent ET0 from increasing. (3) The sensitivity of ET0 to Tmin and Tmax was significantly lower than its sensitivity to other climatic elements. However, increases in temperature can lead to changes in Pwv and Rn. The temperature should be considered the key intrinsic climatic element that has caused the "evaporation paradox" phenomenon in the BTH region.     

Could the Recent Taal Volcano Eruption Trigger an El Ni?o and Lead to Eurasian Warming?

正The Taal Volcano in Luzon is one of the most active and dangerous volcanoes of the Philippines. A recent eruption occurred on 12 January 2020(Fig. 1a), and this volcano is still active with the occurrence of volcanic earthquakes. The eruption has become a deep concern worldwide, not only for its damage on local society, but also for potential hazardous consequences on the Earth's climate and environment.     

Revisiting the Concentration Observations and Source Apportionment of Atmospheric Ammonia

正While China’s Air Pollution Prevention and Control Action Plan on particulate matter since 2013 has reduced sulfate significantly, aerosol ammonium nitrate remains high in East China. As the high nitrate abundances are strongly linked with ammonia, reducing ammonia emissions is becoming increasingly important to improve the air quality of China. Although satellite data provide evidence of substantial increases in atmospheric ammonia concentrations over major agricultural regions, long-term surface observation of ammonia concentrations are sparse. In addition, there is still no consensus on