Characteristics of tropical Pacific SST predictability in coupled GCM forecasts using the NCEP CFS

The limits of predictability of El Niño and the Southern Oscillation (ENSO) in coupled models are investigated based on retrospective forecasts of sea surface temperature (SST) made with the National Centers for Environmental Prediction (NCEP) coupled forecast system (CFS). The influence of initial uncertainties and model errors associated with coupled ENSO dynamics on forecast error growth are discussed. The total forecast error has maximum values in the equatorial Pacific and its growth is a strong function of season irrespective of lead time. The largest growth of systematic error of SST occurs mainly over the equatorial central and eastern Pacific and near the southeastern coast of the Americas associated with ENSO events. After subtracting the systematic error, the root-mean-square error of the retrospective forecast SST anomaly also shows a clear seasonal dependency associated with what is called spring barrier. The predictability with respect to ENSO phase shows that the phase locking of ENSO to the mean annual cycle has an influence on the seasonal dependence of skill, since the growth phase of ENSO events is more predictable than the decay phase. The overall characteristics of predictability in the coupled system are assessed by comparing the forecast error growth and the error growth between two model forecasts whose initial conditions are 1 month apart. For the ensemble mean, there is fast growth of error associated with initial uncertainties, becoming saturated within 2 months. The subsequent error growth follows the slow coupled mode related the model’s incorrect ENSO dynamics. As a result, the Lorenz curve of the ensemble mean NINO3 index does not grow, because the systematic error is identical to the same target month. In contrast, the errors of individual members grow as fast as forecast error due to the large instability of the coupled system. Because the model errors are so systematic, their influence on the forecast skill is investigated by analyzing the erroneous features in a long simulation. For the ENSO forecasts in CFS, a constant phase shift with respect to lead month is clear, using monthly forecast composite data. This feature is related to the typical ENSO behavior produced by the model that, unlike the observations, has a long life cycle with a JJA peak. Therefore, the systematic errors in the long run are reflected in the forecast skill as a major factor limiting predictability after the impact of initial uncertainties fades out.     

Ensemble Forecast： A New Approach to Uncertainty and Predictability

Ensemble techniques have been used to generate daily numerical weather forecasts since the 1990s in numerical centers around the world due to the increase in computation ability. One of the main purposes of numerical ensemble forecasts is to try to assimilate the initial uncertainty （initial error） and the forecast uncertainty （forecast error） by applying either the initial perturbation method or the multi-model/multiphysics method. In fact, the mean of an ensemble forecast offers a better forecast than a deterministic （or control） forecast after a short lead time （3-5 days） for global modelling applications. There is about a 1-2-day improvement in the forecast skill when using an ensemble mean instead of a single forecast for longer lead-time. The skillful forecast （65% and above of an anomaly correlation） could be extended to 8 days （or longer） by present-day ensemble forecast systems. Furthermore, ensemble forecasts can deliver a probabilistic forecast to the users, which is based on the probability density function （PDF） instead of a single-value forecast from a traditional deterministic system. It has long been recognized that the ensemble forecast not only improves our weather forecast predictability but also offers a remarkable forecast for the future uncertainty, such as the relative measure of predictability （RMOP） and probabilistic quantitative precipitation forecast （PQPF）. Not surprisingly, the success of the ensemble forecast and its wide application greatly increase the confidence of model developers and research communities.     

遗传算法优化的BP神经网络在地面温度多模式集成预报的应用研究

基于TIGGE资料集中的ECMWF、CMA和JMA的数值预报产品,利用加权集成、回归集成和消除偏差集成等线性集成方式与遗传算法优化的BP神经网络(GABP)集成,对我国大部开展地面2 m温度在24 h、48 h和72 h预报时效的多模式集成预报试验。通过对2013年1—6月的预报检验,结果表明:GABP集成预报效果有较大提升,均方误差明显小于各单一模式预报。GABP集成的误差分布在新疆和华北均方误差较大,但是在预报效果改进上GABP集成在西部地区相对单一模式的误差减小更加明显。在进行几种多模式集成方式时,GABP集成相比线性方法预报结果更加精准。对于天气过程个例的预报,GABP集成预报出预报量的变化趋势,预报效果优于单一模式和线性集成预报。无论是较长时间段还是短时间的天气过程,在改进预报效果上GABP集成都起到了最佳的作用。     

热带气旋奇异向量在GRAPES全球集合预报中的初步应用

基于副热带奇异向量的初值扰动方法已应用于GRAPES （Global and Regional Assimilation PrEdiction System）全球集合预报系统,但存在热带气旋预报路径离散度不足的问题。通过分析发现,热带气旋附近区域初值扰动结构不合理导致预报集合不能较好地估计热带气旋预报的不确定性,是路径集合离散度不足的可能原因之一。通过建立热带气旋奇异向量求解方案,将热带气旋奇异向量和副热带奇异向量共同线性组合生成初值扰动,以弥补热带气旋区域初值扰动结构不合理这一缺陷,进而改进热带气旋集合预报效果。利用GRAPES全球奇异向量计算方案,以台风中心10个经纬度区域为目标区构建热带气旋奇异向量求解方案,针对台风“榕树”个例进行集合预报试验,并开展批量试验,利用中国中央气象台最优台风路径和中国国家气象信息中心的降水观测资料进行检验,对比分析热带气旋奇异向量结构特征和初值扰动特征,评估热带气旋奇异向量对热带气旋路径集合预报和中国区域24 h累计降水概率预报技巧的影响。结果表明,热带气旋奇异向量具有局地化特征,使用热带气旋奇异向量之后,热带气旋路径离散度增加,路径集合平均预报误差和离散度的关系得到改善,路径集合平均预报误差有所减小,集合成员更好地描述了热带气旋路径的预报不确定性;中国台风降水的小雨、中雨、大雨、暴雨各量级24 h累计降水概率预报技巧均有一定提高。总之,当在初值扰动的生成中考虑热带气旋奇异向量后,可改进热带气旋初值扰动结果,并有助于改善热带气旋路径集合预报效果。      

华南暴雨区域集合预报中不同同化方案的影响试验研究

基于全球集合预报系统(GEFS)资料,利用WRF中尺度模式及GEFS动力降尺度获取区域集合预报初值场,通过对同化后的分析场进行模式积分实现华南前汛期区域集合预报。对2019年6月10日的一次华南前汛期暴雨过程进行不同同化方案的试验:混合同化(Hybrid)、三维变分(3Dvar)、集合卡尔曼滤波(EnKF)和对比试验(Ctrl)四组试验的对比分析,探讨具有不同背景误差协方差矩阵的同化方案对区域集合预报集合扰动和集合离散随时间演变特征的影响,评估不同试验的降水模拟效果。(1) Hybrid对模式初始场有较好的改善作用,而3DVar和EnKF对初始场的改善作用不明显。(2) 对风场、温度场和湿度场,在前期预报中Hybrid的预报误差小于3DVar和EnKF,在中后期的预报中,3DVar和EnKF的预报误差得到改善,且好于Hybrid。同样,集合扰动能量,Hybrid和Ctrl在前期预报发展好于3DVar和EnKF,而在中后期的预报3DVar和EnKF好于Hybrid和Ctrl。(3) 从24 h累积降水评分中,整体上同化试验好于Ctrl,3DVar和EnKF好于Hybrid,且3DVar对大中雨级别的降水评分较好,而EnKF对暴雨以上级别的降水评分较好。(4) 对于集合统计检验分析,同化试验的AUC值都大于Ctrl的AUC值,24 h累积降水量阈值在10~100 mm的AUC值,3DVar最好;而125 mm阈值的AUC值,EnKF最好。      

基于动力降尺度的区域集合预报初值扰动构建方法研究

利用全球集合预报系统资料(Global Ensemble Forecast System,GEFS),基于WRF中尺度模式构建了区域集合预报系统,区域集合初值的构建采用两种方案,一种是GEFS全球集合预报初值场直接动力降尺度(称为DOWN集合),另一种是提取GEFS全球集合降尺度后的扰动场,并叠加到区域数值预报系统(北京快速更新循环数值预报系统:Beijing Rapid Update Cycle System,BJ-RUC)分析场上构建集合初值场(称为D-RUC集合)。进行了批量试验,通过对比发现D-RUC集合的中小尺度扰动增长优于DOWN集合,而大尺度扰动分量的增长两者相当,说明与高分辨率分析场叠加可以促进动力降尺度扰动的中小尺度扰动分量的增长。集合预报扰动准确性检验结果显示,短预报时效内DOWN集合扰动明显低估了预报误差,在预报误差较大的位置扰动较小,而D-RUC集合能够更好地识别预报场中哪些位置预报误差较大,而哪些位置预报误差较小。集合预报检验结果表明,D-RUC方法能显著改善短时效预报效果,集合离散度有所增加、均方根误差有所减少,概率预报评分显示D-RUC集合比DOWN集合在短预报时效占优。降水个例分析结果表明D-RUC方法能显著改善短时效内的降水概率预报效果。     

Impact of Perturbation Schemes on the Ensemble Prediction in a Coupled Lorenz Model

Based on a simple coupled Lorenz model, we investigate how to assess a suitable initial perturbation scheme for ensemble forecasting in a multiscale system involving slow dynamics and fast dynamics. Four initial perturbation approaches are used in the ensemble forecasting experiments: the random perturbation(RP), the bred vector(BV), the ensemble transform Kalman filter(ETKF), and the nonlinear local Lyapunov vector(NLLV) methods. Results show that,regardless of the method used, the ensemble ave...     

Evaluation of Tropical Cyclone Intensity Forecasts from Five Global EnsemblePrediction Systems During 2015-2019

This study presented an evaluation of tropical cyclone (TC) intensity forecasts from five global ensemble prediction systems (EPSs) during 2015-2019 in the western North Pacific region. Notable error features include the underestimation of the TC intensity by ensemble mean forecast and the under-dispersion of the probability forecasts.The root mean square errors (brier scores) of the ensemble mean (probability forecasts) generally decrease consecutively at long lead times during the five years, but fluctuate between certain values at short lead times.Positive forecast skill appeared in the most recent two years (2018-2019) at 120 h or later as compared with the climatology forecasts. However, there is no obvious improvement for the intensity change forecasts during the 5-yearperiod, with abrupt intensity change remaining a big challenge. The probability forecasts show no skill for strongTCs at all the lead times. Among the five EPSs, ECMWF-EPS ranks the best for the intensity forecast, while NCEP-GEFS ranks the best for the intensity change forecast, according to the evaluation for ensemble mean and dispersion. As for the other probability forecast evaluation, ECMWF-EPS ranks the best at lead times shorter than 72 h, while NCEP-GEFS ranks the best later on.     

GRAPES区域集合预报条件性台风涡旋重定位方法研究

为了在集合预报中更合理描述台风涡旋中心定位的不确定性,采用2009—2018年中国气象局和日本气象厅台风最佳路径数据,分析台风最佳路径涡旋中心定位的不确定性特征,在此基础上设计条件性台风涡旋重定位方法（Conditional Typhoon Vortex Relocation,CTVR）,构建集合成员台风涡旋中心重定位阈值条件、台风涡旋分离数学处理及涡旋重定位等数学处理过程,利用中国气象局数值预报中心区域集合预报系统（Global/Regional Assimilation and Prediciton System-Regional Ensemble System,GRAPES-REPS）对2018年西北太平洋上的3个台风（1808号“玛莉亚”、1824号“谭美”和1825号“康妮”）进行轴对称结构和轴对称+非对称结构条件性台风涡旋重定位两种方案的集合预报试验和检验评估。结果表明:（1）中国气象局和日本气象厅台风最佳路径误差平均值为13.72 km,可视为台风涡旋中心定位不确定性的合理估计值;（2）统计检验结果和典型个例分析表明,采用轴对称结构和轴对称+非对称结构条件性台风涡旋重定位方法的台风集合预报路径误差及集合预报一致性结果比较接近;（3）条件性台风涡旋重定位方法可以有效改进GRAPES-REPS区域集合预报台风路径概率预报效果,如台风路径集合预报平均误差有所减小,集合预报一致性（路径离散度与路径均方根误差比值）增大,特别是预报初期概率预报效果改进更为显著,而预报中后期改进有限;（4）通过对“玛莉亚”台风集合预报诊断分析发现,经过条件性台风涡旋重定位后,各集合成员的台风路径误差在预报初期明显减小且路径收敛,但随着预报时效的延长台风路径逐渐发散。应用条件性台风涡旋重定位方法后,台风涡旋环流与大尺度环境场仍然比较连续协调,且台风涡旋环流外的大尺度环境场具有一致性特点,最低气压误差、最大风速误差和降水预报技巧基本不变。可见,条件性台风涡旋重定位方法的应用可以提供更准确的台风路径预报不确定性信息,帮助预报员做出更准确的预报决策。      

GRAPES区域集合预报模式的初值扰动增长特征

基于GRAPES-REPS（Global and Regional Assimilation and Prediction Enhanced System-Regional Ensemble Prediction System）区域集合预报模式和集合变换卡尔曼滤波（Ensemble Transform Kalman Filter,ETKF）初值扰动方法,对2015年6月1~15日10 km与15 km水平分辨率分别进行集合预报试验,通过分析ETKF初值扰动分量、初值扰动方差准确率、动能谱、扰动能量演变、日变化及集合离散度、均方根误差等特征,揭示GRAPES-REPS区域集合预报ETKF初值扰动结构及增长特征。结果表明:（1）ETKF初值扰动方案产生的扰动能够保持所有正交、不相关方向的误差方差,且ETKF分量α参数值及放大因子具有较好的稳定性。（2）ETKF初值扰动方法生成的扰动场以大尺度扰动为主,扰动结构及能量具有随流型依赖特征,低层以内能扰动为主,高层以动能扰动为主,且集合扰动可以有效捕捉预报误差的结构。（3）GRAPES区域集合预报初值扰动总能量和集合离散度随预报时效的延长均呈发展趋势,但离散度增长率小于均方根误差增长率,即集合预报总体存在集合离散度不足的问题。（4）水平分辨率提高可以增加中高层大尺度扰动波谱能量,明显改进等压面及近地面风场及温度场的集合预报效果。值得指出的是,GRAPES-REPS区域集合预报低层内能扰动能量存在明显的日变化特征,特别是青藏高原地区更加显著,需要进一步研究青藏高原初值扰动结构的合理性。     

多物理ETKF在暴雨集合预报中的初步应用

基于集合转换卡尔曼滤波（ETKF）的初值扰动方法是目前集合预报领域热点方法之一,但应用在短期集合预报中仍存在离散度不够、误差较大等问题。考虑到在区域短期集合预报中,模式不确定性和边界不确定性的影响不能忽略,本文尝试在ETKF生成分析扰动的过程中,同时考虑初值不确定性、物理不确定性与边界不确定性,进而构建多初值、多物理、多边界ETKF集合,并以2010年9月30日到10月8日海南岛特大暴雨作为研究个例,对其在暴雨集合预报中的应用展开初步研究,重点分析多种物理参数化过程对预报结果的影响。结果表明,多物理过程的ETKF（多物理ETKF）和单物理过程的ETKF（单一ETKF）均优于对照预报,多物理ETKF优势更加明显,其均方根误差、离散度等指标均得到很好的改善;对于降水采用SAL方法进行检验,发现多物理ETKF对于降水位置的预报有明显的改善,对于特大暴雨的强度预报也略有改善。研究表明,在ETKF初值扰动中加入多种物理过程,可以有效改善短期集合的离散度,提高预报准确率,有良好的发展前景和应用潜力。     

集合变换卡尔曼滤波局地化对区域集合初始扰动的影响

集合变换卡尔曼滤波(ensemble transform Kalman filter, ETKF)是一种有效的集合预报初始扰动构造方案。但是,有限的集合样本、相同的集合成员设置以及预报模式误差等可能会使两个距离较远的状态变量产生虚假相关,从而影响ETKF集合扰动的质量。为了有效解决远距离虚假相关问题,将局地化思想引入ETKF方案。本文针对GRAPES区域集合预报系统(GRAPES REPS),对ETKF初值扰动局地化方案的效果进行了试验分析,为进一步改善和优化局地化方案(LETKF方案)提供依据。通过一周的连续试验,从暴雨个例、集合预报多种评分检验等方面分析了LETKF初始扰动方案所产生的集合预报质量。结果表明,区域集合预报中集合变换卡尔曼滤波初始扰动的局地化方案能够更加合理地捕捉到快速增长的分析误差的物理结构,更准确地再现数值模式预报误差的线性与非线性传播和演变特征。该局地化方案可以较好地改进预报质量,提高降水预报的准确率,尤其是针对小雨、中雨、暴雨量级的预报。相对于现有区域集合预报的业务系统GRAPES REPS,基于局地化ETKF初始扰动方案的区域集合预报具有较明显的优势。总体来看,LETKF初始扰动方案可更好地改善区域集合预报的质量。     

Forecasts of hydrological parameters over the Mackenzie River Basin: Sensitivity to initial conditions,horizontal resolution and forecast range

Abstract

As part of the Global Energy and Water Cycle Experiment, Canadian global spectral forecast model predictions of surface water and energy fluxes over the Mackenzie River basin are examined. Two nine‐member ensemble forecasts of one month duration are produced with the operational model, for a spring and a summer case, at a horizontal resolution of about 100 km (T95). The sensitivity to initial conditions is measured by the degree to which the individual forecasts in the ensembles vary one from another. The evolution in time of this estimated error (ensemble standard deviation) is determined for the surface energy and water accumulations, averaged over the basin. For comparison the calculations are repeated for the Mississippi basin and over North America. The greatest sensitivity is found for the net accumulation of precipitation minus evaporation. The spring ensemble is redone at a coarser horizontal resolution (T47), and the results are similar. The forecast uncertainty (ensemble standard deviation) of the area‐averages over the basin appear to be unaffected by this change, although the ensemble mean values are sensitive to the change in resolution. The ensemble standard deviation makes a significant, abrupt increase toward the end of the second week into the forecasts. This investigation suggests a need for an improved model, if the forecasts’ useful range is to extend to one month. Available upgrades to the land‐surface, precipitation and evaporation schemes will be used in subsequent work, and the forecasts reported here will serve as a baseline for comparison.     

Study on ETKF-Based Initial Perturbation Scheme for GRAPES Global Ensemble Prediction

Initial perturbation scheme is one of the important problems for ensemble prediction. In this paper, ensemble initial perturbation scheme for Global/Regional Assimilation and PrEdiction System (GRAPES) global ensemble prediction is developed in terms of the ensemble transform Kalman filter (ETKF) method.A new GRAPES global ensemble prediction system (GEPS) is also constructed. The spherical simplex 14-member ensemble prediction experiments, using the simulated observation network and error characteristics of simulated observations and innovation-based in ation, are carried out for about two months. The structure characters and perturbation amplitudes of the ETKF initial perturbations and the perturbation growth characters are analyzed, and their qualities and abilities for the ensemble initial perturbations are given. The preliminary experimental results indicate that the ETKF-based GRAPES ensemble initial perturbations could identify main normal structures of analysis error variance and reflect the perturbation amplitudes.The initial perturbations and the spread are reasonable. The initial perturbation variance, which is approximately equal to the forecast error variance, is found to respond to changes in the observational spatial variations with simulated observational network density. The perturbations generated through the simplex method are also shown to exhibit a very high degree of consistency between initial analysis and short-range forecast perturbations. The appropriate growth and spread of ensemble perturbations can be maintained up to 96-h lead time. The statistical results for 52-day ensemble forecasts show that the forecast scores ofensemble average for the Northern Hemisphere are higher than that of the control forecast. Provided that using more ensemble members, a real-time observational network and a more appropriate inflation factor,better effects of the ETKF-based initial scheme should be shown.     

基于BGM的不同繁殖长度对集合预报的影响

利用T63L9全球谱模式和NCEP/NCAR再分析资料, 对BGM方法中增长模的繁殖长度对集合预报效果的影响进行研究。结果表明:与控制预报相比, 不同繁殖长度的集合预报都能使预报效果得到一定程度的改进, 特别是第4天预报以后, 改进程度随预报时效而稳步提高。三组不同繁殖长度的集合预报对控制预报的改进存在差别, 分析结果表明:繁殖长度为2 d的集合预报明显效果最差, 而繁殖3 d和4 d的集合预报差别并不明显。对集合Talagrand分布以及离散度的初步分析表明, 繁殖长度取为3 d似乎最为合理。     

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。     

集合平均方法减小混沌系统计算误差的效果研究

研究了Lorenz非线性系统中使用的集合平均方法来减小计算误差的效果,通过检查5组数值试验(每组20个样本)的结果发现:集合平均对计算误差的减小和消除不如高精度算法有效,这主要体现在以下几方面:1)普通的算法和双精度的计算环境中,若截断误差是主导误差(当初值误差很小时),各集合的平均结果并不收敛于真值,而是收敛于含截断误差的数值解;2)若初值误差为主导时,系统受到初值误差增长规律的影响,数值解收敛于由初值误差主导的误差解;3)这两种误差量级接近的时候,两种误差都无法消除掉。对解的统计特征进行研究表明,可信的数值解与含计算误差的数值解有许多相似的地方,但是与集合平均的数值解有很大不同,同样说明了集合平均不适用于减小计算误差这样的问题。此外,试验结果表明即使数值解的概率分布形式基本正确,也不能保证数值解是正确的。     

基于GRAPES_Meso的集合预报扰动方案设计与比较

基于GRAPES_Meso区域集合预报系统,设计了三种集合预报扰动方案,即多初值、多初值多物理、多初值多物理多边值,并针对三种方案进行了连续一个月的批量试验,重点分析了2008年7月23日江淮暴雨过程.结果表明,对于降水预报,三种集合扰动方案均相对于控制预报均有所改善,多初值多物理与多初值多物理多边值方案对小雨、中雨预报改进效果显著,对暴雨预报略有改进;多初值方案仅能产生有限的集合离散度且难以增长,引入物理参数方案扰动及边界条件扰动能显著提高集合离散度,改善各物理量场的预报效果;通过比较,多初值多物理多边值为最优方案.该批量试验表明,模式物理过程及边界条件是影响GRAPES _Meso区域集合预报不确定性的不可忽视因素.     

集合预报误差在GRAPES全球四维变分同化中的应用研究Ⅱ：参数选取与数值试验

研究的第一部分讨论了如何有效应用集合预报误差的科学方案,确定了集合预报误差在GRAPES（Global Regional Assimilation and PrEdiction System）全球4DVar（four dimensional variational data assimilation）中应用的分析框架。在此基础上研究了针对集合预报误差实际应用于GRAPES全球4DVar,解决接近或超过100个集合样本数时高效生成的计算效率问题,以及与GRAPES全球4DVar匹配的同化关键参数确定问题。选择基于4DVar的集合资料同化方法生成集合样本,通过将第1个样本极小化迭代过程中产生的预调节信息用于其他样本极小化做预调节,将计算效率提高了2倍。通过时间错位扰动方法增加集合样本数,实现集合样本增加到3倍。对集合方差进行膨胀,并选择水平局地化相关尺度为流函数背景误差水平相关的1.4倍。通过批量数值试验方法确定背景误差与集合预报误差的权重系数,对60个集合样本当集合预报误差权重为0.7时预报效果最好。对北半球夏、冬两季各52 d的批量试验表明,对于南、北半球En4DVar （ensemble 4DVar）较4DVar的改进在冬季主要集中在700—30 hPa,而在夏季主要集中在400—150 hPa。赤道地区受季节影响较小,En4DVar对位势高度、风场与温度的改进都较为明显,且经向风场的改进最为显著。文中研发的集合预报误差在GRAPES全球4DVar中应用的方法合理可行。      

基于KPCA的台风强度神经网络集合预报方法研究

针对影响台风强度前期预报因子较多以及因子的非线性变化特点,首先采用逐步回归方法筛选出部分预报因子,再利用核主成分分析方法在剩余的预报因子中提取包含了原数据较多信息的核主成分与前期选入的预报因子共同作为模型输入.进一步考虑到神经网络集合预报中个体的准确性和差异性的权衡问题,在不同的初始条件下生成若干组神经网络,分别选择每组中性能最优的个体,建立了一种新的非线性神经网络集合预报模型.最后以西北太平洋海域2001-2010年5-10月的台风强度为研究对象进行了预报试验.结果表明,这种神经网络集合预报模型的预报结果符合实际应用的要求,其预报平均绝对误差明显小于同等条件下的神经网络方法和逐步回归预报方法.