基于人工智能大模型的上海2024年极端高温事件次季节预测
doi: 10.13878/j.cnki.dqkxxb.20250109002
梁萍1 , 张志琦1 , 曹欣沛2 , 黄文娟1
1. 上海市气候中心/中国气象局上海城市气候变化应对重点开放实验室/上海市气象与健康重点实验室,上海 200030
2. 成都信息工程大学大气科学学院,四川 成都 610225
基金项目: 国家自然科学基金项目(U2342208) ; 国家重点研发计划项目(2024YFC3013100) ; 上海市自然科学基金项目(24ZR1492500) ; 中国气象局复盘总结项目(FPZJ2025-043) ; 中国气象局重点创新团队项目(CMA2023ZD03) ; 中国气象局青年创新团队项目(CMA2024QN06)
Sub-seasonal prediction of extreme heatwave events in Shanghai for 2024 using artificial intelligence-driven large models
LIANG Ping1 , ZHANG Zhiqi1 , CAO Xinpei2 , HUANG Wenjuan1
1. Shanghai Regional Climate Center/Key Laboratory of Cities' Mitigation and Adaptation to Climate Change in Shanghai,China Meteorological Administration/Shanghai Key Laboratory of Meteorology and Health,Shanghai 200030 ,China
2. School of Atmospheric Sciences,Chengdu University of Information Technology,Chengdu 610225 ,China
摘要
近年来高温事件频发、强发,在此背景下提高高温预测的准确性显得尤为重要。目前气象人工智能大模型发展迅速,但其对超大城市极端高温的次季节预测效果尚不清楚。本文以2024年盛夏上海极端高温事件为例,采用3个人工智能气象大模型(Pangu、FuXi和FourCastNet)的预测数据,利用相关技巧、功率谱分析等方法,对比评估了气象大模型对该年高温及其环流的预测效果。结果表明,上海盛夏高温日数与其上空副热带高压强度呈显著的正相关关系,2024年盛夏上海的持续性高温时段与其10~20 d准周期振荡的正位相基本一致,同时受到30~60 d低频振荡的调制。在3个大模型中,Pangu、FuXi在提前15 d内可提供有技巧的高温预测参考,部分大模型(Pangu)对影响高温的副热带高压演变的有效超前预测时效可达16~20 d。不同大模型的次季节预测效果差异与其对高温及其关键环流系统的低频演变的预测能力有关。大模型对影响上海的低频环流演变的预测能力越强,对副热带高压及高温的次季节预测效果就越好。相较于出梅和台风影响阶段,大模型在该年盛夏的其他时段可提供更大的次季节高温预测机会窗口。
Abstract
With the intensification of global warming,extreme heatwave events are occurring with increasing frequency and intensity,posing severe impacts to human society and ecosystems.Accurate prediction of extreme heatwaves is essential for urban resilience,as megacities are particularly vulnerable due to impacts on public health,energy supply,and transportation.Sub-seasonal prediction,which bridges short-term weather forecasting and seasonal prediction,plays a critical role in mitigating the effects of extreme heatwaves.However,traditional sub-seasonal prediction studies have primarily focused on temperature anomalies or their low-frequency components.Recent advancements in meteorological artificial intelligence (AI) models have led to the rapid developments in weather prediction,but their capability for sub-seasonal heatwave prediction in megacities remains unclear.This study evaluates the performance of three AI-driven meteorological models (Pangu,FuXi,and FourCastNet) in predicting heatwave events in Shanghai during midsummer 2024.The evaluation is based on correlation skill,power spectrum analysis,and other diagnostic methods.Analysis of heatwaves and their associated circulation patterns indicates that the number of high-temperature days in Shanghai during midsummer is significantly and positively correlated with the strength of the subtropical high aloft.The timing of heatwave occurrences aligns closely with the positive phases of the 10—20 d quasi-periodic oscillation and is simultaneously influenced by the 30—60 d low-frequency oscillation.On the quasi-biweekly timescale,extreme heatwave events in Shanghai are linked to both the northwestward propagation of wave trains triggered by convective anomalies in the tropical western Pacific and the eastward propagation of circulation anomalies along the Silk Road teleconnection pattern in the middle and high latitudes.Among the three AI models,both Pangu and FuXi demonstrate skillful high-temperature predictions up to 15 d in advance.Certain models,such as Pangu,can effectively predict the evolution of the subtropical high associated with heatwaves at lead times of 16—20 d.Differences in sub-seasonal prediction performance among the AI models are associated with their ability to capture low-frequency evolution of high temperatures and atmospheric circulation.Models with stronger predictive capabilities for low-frequency circulation anomalies tend to produce more accurate sub-seasonal forecasts of the subtropical high and high-temperature events.Furthermore,AI models exhibit greater predictive skill for sub-seasonal heatwaves during periods not affected by the end of Meiyu season or typhoon activity.This study also highlights that the differences in sub-seasonal prediction skill among the three AI models for low-frequency circulation at lead times of 11—15 d are influenced by the intensity of the initial low-frequency state.The uncertainty in prediction skill caused by initial-state variability warrants further investigation.Given that this study evaluates AI model performance primarily using 2024 as a case study,more comprehensive assessments are needed.Additionally,when applying AI models to enhance sub-seasonal heatwave prediction in megacities,the influence of surface heterogeneity,such as urbanization effects,should be considered to improve forecast accuracy.
近年来,在全球气候变暖加剧的背景下,极端高温事件在全球范围内频繁发生,对人类社会和自然生态系统造成了严重影响(IPCC,2021),特别是在包括中国在内的亚欧大陆(Rogers et al.,2022)。超大城市由于其人口密集、建筑密集和人为热排放等因素,更容易受到极端高温的威胁(Oke,1982; Grimmond et al.,2010; Liang et al.,20222024; Ma et al.,2022)。极端高温对人体健康(Basu,2009; Gasparrini et al.,2015)、能源供应、交通和基础设施等带来巨大挑战(Liu et al.,2024)。准确预测超大城市的极端高温事件是城市安全运行和可持续发展的重要需求。
次季节预测作为一种介于短期天气预报和季节气候预测之间的时间尺度预测,对于极端高温事件的应对具有重要意义(Vitart et al.,2017),是近年来和未来10年的重要国际研究方向(NASEM,2016; WMO/WWRP,2024)。然而,次季节预测面临着诸多挑战,例如大气系统的复杂性和不确定性、初始条件的敏感性,以及不同物理过程之间的相互作用等(Vitart et al.,2017; Liang and Lin,2018)。基于S2S(sub-seasonal to seasonal prediction)计划的动力模式开展的气温次季节预测评估表明,动力模式对中国东部气温异常的次季节预测有效技巧可达到3周甚至在部分区域可达4周(Liang and Lin,2018)。东亚夏季地表气温存在显著的次季节变率(Liang et al.,2018),是次季节预测的一个重要来源。Zhu and Li(2018)杨秋明(2018)雷蕾等(2022)Xie et al.(2024)根据气温演变及与其相联系的前期低频关键影响信号,采用统计回归、卷积神经网络等方法开展了长江流域夏季高温低频分量的次季节预测模型构建研究,有技巧预测时效可提前20~30 d。而基于S2S动力模式对夏季气温次季节变率主模态的有效预测时长可达3周,优于动力模式直接输出的2周(Liang et al.,2018)。然而,已有研究多针对气温异常或者气温的低频分量开展预测,针对极端高温事件的次季节预测仍存在局限性。即使是最先进的统计次季节模型,在提前15 d的情况下仅能捕捉到中国近30%的热浪(Zhu and Li,2018)。部分时间段(如7月下旬)长江流域日最高温度的次季节动力预报技巧存在显著的次季节预报技巧障碍,这与气候态季节内变化的阶段性转折影响有关(Yang et al.,2018)。
近年来,人工智能技术的快速发展为气象预测带来了新的机遇和方法(Goodfellow et al.,2016)。人工智能大模型凭借其强大的学习能力和对复杂数据的处理能力,在气象领域展现出了巨大的应用潜力(Molina et al.,2023)。目前主流人工智能气象大模型有FourCastNet(Pathak et al.,2022)、GraphCast(Lam et al.,2023)、Pangu(Bi et al.,2023)、FuXi(Chen K et al.,2023)、FengWu(Chen L et al.,2023)等,对中期气象要素和极端降水、强风暴等高影响事件的预报具有预报时效长、精度高、预测推理快的特点(Charlton-Perez et al.,2024; 黄小猛等,2024; Olivetti and Messori,2024; Xu et al.,2024)。近期,中国气象局针对中短期预报、临近预报和次季节-季节预测分别发布了人工智能气象大模型系统“风清”“风雷”“风顺”(https ://www .cma .gov .cn/2011xwzx/2011xqxxw/2011xqxyw/202406/t20240618_6359467.html)。然而,将人工智能大模型应用于超大城市极端高温事件的次季节预测仍处于起步阶段,需要进一步探索和深入研究。2024年夏季,包括上海在内的长江流域出现持续性高温,其中上海市区高温(日最高气温≥35℃)日数达52 d(在近150年的观测记录中排名第二,仅次于1934年的55 d)。为此,本文以2024年上海盛夏为例,评估3个气象大模型Pangu、FuXi和FourCastNet对上海盛夏高温及相关环流形势的预测结果,为基于人工智能气象大模型开展城市极端高温的次季节预测提供思路和参考。
1 资料和方法
资料包括:1)ECMWF第五代再分析数据集ERA5(Hersbach et al.,2020),时间分辨率为4次/d,水平分辨率为31 km,垂直方向从地表至80 km共137层。2)美国大气海洋局开发的全球逐日向外长波辐射(OLR)资料(Liebmann and Smith,1996)。3)由中国气象局国家气象信息中心开发的1979—2024年CRA-40逐日网格化地表气温再分析数据(Liang et al.,2020),水平分辨率为0.5°×0.5°。4)来源于上海市气象信息中心的上海11个基本气象站自1961年以来的气温观测资料。
利用3个数据驱动的人工智能气象大模型(FourCastNet、Pangu、FuXi)来开展高温和环流形势的预测。其中,使用ERA5再分析数据集作为机器学习预测模型的初始输入数据和预测评估数据。预训练模型使用ECMWF开发的AI模型工具箱(https ://github .com/ecmwf-lab/ai-models)生成。上述3个大模型同属自回归模型,可从相同初始化分析场出发,给定时间步长逐步输出可用于预测下一个时间步长的变量。3个大模型的具体介绍可参考Pathak et al.(2022)、Chen L et al.(2023)和Bi et al.(2023)。为了减少样本少的影响,将每隔5 d的超前预测结果归为一类评估。如7月10日的超前1~5 d的预报结果,是7月5—9日分别超前5~1 d对7月10日的预报结果; 7月20日的超前16~20 d的预报结果,是6月30日—7月4日分别超前20~16 d对7月20日的预测结果,其他依此类推。为进一步了解大模型对低频振荡预测效果评估,本文选取欧洲中期天气预报中心(ECMWF)的S2S模式(简称EC)预测结果与大模型预测进行比较,数据来自中国气象局的S2S数据中心(http ://s2s .cma .cn/index)。
在气温演变诊断和预报技巧评估中,主要采用集合经验模态分解(EEMD)方法(Wu and Huang,2009)获得其不同时间尺度的波动分量。该方法的最大优点在于能够以自适应方式提取信号的不同分量。
在开展人工智能大模型的预测效果评估中,将ERA5逐日14时再分析气温资料视作日最高气温实况值与预测值进行比较。其中,通过2008—2023年上海全市平均日最高气温观测值与盛夏逐日14时气温再分析资料的对比发现,两者相关显著(相关系数为0.95),即再分析资料能反映最高气温的逐日变化。但需注意的是,再分析资料较观测值有明显的系统弱偏差。
2 2024年盛夏上海高温及环流背景
图1a给出了2024年盛夏(7—8月)东亚气温距平的空间分布。可见,该年盛夏中国东部平均气温均较常年偏高,其中以长江流域最为明显,包括上海在内的长三角地区平均气温偏高2℃以上。由1961—2024年上海全市11个基本站平均气温演变(图1b)可见,2024年盛夏平均气温较1991—2020年气候平均偏高2.6℃,为1961年以来历史记录次高(仅次于2013年),其中徐家汇、浦东、闵行、惠南站分别列1961年以来最高(图略)。此外,上海市区徐家汇2024年盛夏高温日数为47 d,在1873年以来的气象记录中排名第二(仅次于1934年的48 d)。
12024年盛夏(7—8月)东亚气温距平的空间分布(a; 单位:℃)和1961—2024年上海全市盛夏平均气温的演变(b; 单位:℃)
Fig.1(a) Spatial distribution of temperature anomalies (units:℃) in East Asia during midsummer (July-August) 2024, and (b) evolution of the average temperature (units:℃) in Shanghai during midsummer from 1961 to 2024
由2024年盛夏500 hPa环流异常(图2a)可见,西太平洋副热带高压较常年偏强、偏西、偏大,导致长三角地区长期受副热带高压下沉气流控制。进一步从上海上空(121°~122°E,30.5°~31.5°N,下同)500 hPa位势高度和上海高温日数的逐年演变(图2b)可见,去趋势(原始逐年序列减去其线性拟合项)后的500 hPa位势高度与盛夏上海高温日数呈显著的正相关关系,1991—2023年的相关系数达0.63(通过0.001信度的显著性检验)。2024年盛夏期间影响上海上空的高压强度列1961年以来第三位(仅次于2022和2010年),是同期上海出现极端高温事件的重要环流背景。
利用集合经验模态分解方法获得上海全市平均日最高气温及其上空500 hPa位势高度在不同时间尺度上的变化分量,除天气尺度波动和季节循环外,包括两个季节内低频分量,分别对应10~20 d低频振荡(LFO1)和30~60 d低频振荡(LFO2)。由图3a可见,上海高温(全市平均的日最高气温≥35℃)过程出现在7月上旬中后期、中旬中后期至下旬前期、8月上半月以及8月下旬前中期,与准双周振荡(LFO1)的正位相时段一致率高达94%(8月上旬中后期除外)。同时,上海高温过程还受到30~60 d低频振荡(LFO2)的调制,除对高温过程的强度有增幅作用外,对8月上旬中后期和下旬前中期高温的持续出现有重要影响。类似地,上海上空500 hPa位势高度演变反映出副热带高压在7月上旬中后期、中旬中后期至下旬前期、7月下旬后期至8月上旬以及8月中旬后期至下旬前中期呈现阶段性增强过程,这也主要与其准双周振荡(LFO1)低频分量有关,并受到30~60 d低频振荡(LFO2)的调制影响(图3b)。
22024年盛夏(7—8月)500 hPa位势高度异常(a; 阴影,单位:gpm; 细(粗)实线、细(粗)虚线分别表示5 760(5 880)gpm等值线及与之对应的气候平均值(1991—2020年平均)),以及去趋势(原始逐年序列减去其线性拟合项)的1961—2024年上海全市平均盛夏高温日数异常(红线,单位:d)及其上空(121°~122°E,30.5°~31.5°N)500 hPa位势高度异常(蓝线,单位:gpm)的演变(b)
Fig.2(a) Anomalies of the500 hPa geopotential height during midsummer (July-August) 2024 (shadings, units:gpm; thin (thick) solid and dashed lines represent the5 760 gpm (5 880 gpm) contours and the corresponding climatology from 1991 to 2020) , and (b) detrended evolution of anomalies in the average number of high-temperature days (red curve, units:d) in Shanghai and 500 hPa geopotential height (blue curve; units:gpm) over Shanghai (121°—122°E, 30.5°—31.5°N) during midsummer from 1961 to 2024
32024年盛夏(7—8月)上海全市平均日最高气温Tmax(a; 单位:℃)及其上空500 hPa位势高度z500(b; 单位:dagpm)的逐日演变及其低频分量(LFO1、LFO2分别表示10~20 d、30~60 d低频分量)
Fig.3Daily evolution and its low-frequency components of (a) average daily maximum temperature (units:℃) in Shanghai and (b) 500 hPa geopotential height (units:dagpm) over Shanghai during midsummer (July-August) 2024 (LFO1 and LFO2 represent low-frequency components with 10—20 d and 30—60 d periods, respectively)
图4进一步给出了与上海上空500 hPa位势高度及Tmax准双周振荡相关联的前期至同期500 hPa位势高度场和OLR准双周低频演变。由图4a—e可见,超前12 d左右由西北大西洋经欧亚大陆东传至长江下游的丝绸之路遥相关波列,在超前0 d显著影响上海上空500 hPa位势高度的准双周振荡异常,导致控制上海上空的副热带高压显著增强。同时,在准双周尺度上,赤道南北两侧的西太平洋热带异常对激发的波列自超前12 d左右开始向西北方向传播,在超前0 d包括上海在内的长江下游对流出现显著负异常,可进一步通过局地的云覆盖效应以及与环流异常相联的下沉增温,促进高温的增强(图4f—j)。由此可见,在准双周时间尺度上,热带西太平洋对流异常激发波列的西北向传播,以及中高纬环流沿丝绸之路遥相关波列的东传,通过影响上海上空500 hPa位势高度异常及热辐射异常等,促进极端高温事件的形成。
3 大模型对高温的预测效果评估
从3个大模型超前不同时效对上海日最高气温的预测与实况(ERA5再分析资料)的对比(图5)来看,提前10 d的预测基本能反映高温的实况演变。其中,3个大模型提前1~5 d的预测效果差别不明显(与实况的相关系数为0.91~0.93); 提前6~10 d的预测与实况虽然相关系数(0.66~0.83)有一定差异,但均通过了0.001信度的显著性检验。在次季节尺度上,3个大模型提前11~15 d的高温预测效果差别较大。例如,Pangu、FuXi、FourCastNet提前11~15 d的高温预测与实况的相关系数分别为0.73、0.27和0.04,Pangu和FuXi分别通过了0.001和0.05信度的显著性检验。随着超前时效的延长,预测性能明显减弱,对于可提供超前16~20 d预测的Pangu而言,其预测与实况的相关系数降为0.11。这表明,不同的大模型对2024年盛夏上海高温的次季节预测效果不同,部分模型在提前15 d内可提供有技巧的预测参考。
大模型在次季节尺度(如提前11~15 d)对高温的预测效果差异与其对高温低频演变的反映能力有关。对比实况与大模型超前11~15 d预测的最高气温的功率谱(图6)可见,Pangu预测的最高气温演变功率谱与实况接近,均能反映出显著的10~20 d低频振荡的特点,且50~60 d低频振荡也较为明显; FourCastNet的预测尽管能反映高温演变的10~15 d低频振荡特点,但其反映出的相对明显的20~40 d低频振荡与实况相反; FuXi的预测效果则介于Pangu和FourCastNet之间。由于上海高温过程与其低频演变特别是10~20 d低频活动密切相关(图3a),故3个大模型对高温低频演变的反映能力直接影响高温过程的预测效果。这也进一步解释了3个大模型对该年上海高温次季节预测效果存在差异的原因。
另一方面,对比模型不同阶段不同超前时效的预测误差(图7)可见,除所有模型对高温阶段的预测存在弱偏差的共同特点外,随着预测时效的延长,模型预测与实况的差异均有所增大,但不同阶段的预测误差存在差异。总体上,所有模型对7月的高温预测误差大于8月,特别是7月上旬后期至中旬前期和7月下旬后期的误差更为明显。其中,7月上旬后期至中旬前期为西太平洋副热带高压出现异常的阶段性南落时期,对应长江中下游从断梅形势转为二段梅形势; 7月下旬后期为台风“格美”影响长江中下游时段。由此可见,在长江中下游常年出梅阶段,大模型的预测误差较大,这与Liang and Lin(2018)针对S2S计划的次季节业务模式的预测类似。同时,大模型对台风影响长江中下游阶段的上海日高气温预测的误差也较大,这与Yang et al.(2018)关于S2S计划的次季节模式存在7月末的高温预测障碍结论一致。这表明,对2024年盛夏高温预测而言,出梅和台风影响阶段之外的大模型次季节预测效果更好,可提供更大的机会窗口。
4上海上空500 hPa位势高度的10~20 d低频分量与超前12 d(a)、9 d(b)、6 d(c)、3 d(d)和0 d(e)的10~20 d低频滤波后的500 hPa位势高度距平场的回归系数分布(单位:gpm; 交叉影线区域表示通过0.1信度的显著性检验);(f—j)同(a—e),但为Tmax低频分量与OLR的回归系数分布(单位:W/m2
Fig.4Spatial distribution of regression coefficients between the10—20 d low-frequency component of the500 hPa geopotential height over Shanghai and the anomaly fields of the10—20 d low-pass filtered 500 hPa geopotential height at lead times of (a) 12 d, (b) 9 d, (c) 6 d, (d) 3 d, and (e) 0 d (units:gpm; cross-hatched areas indicate statistically significant coefficients at the 0.1 level) . (f—j) as in (a—e) , but for the regression coefficients between the low-frequency component of Tmax and OLR (units:W/m2)
由前述可知,大模型对2024年盛夏上海高温的有技巧预测时效最长可达11~15 d。考虑到上海高温的出现与其上空副热带高压的影响密切相联,本文进一步考察大模型对上海上空对流层中层位势高度的预测效果。由500 hPa位势高度的实况演变(图8k)可见,影响上海上空的副热带高压在7月上旬前期、7月中旬中期至下旬前期、7月下旬后期至8月上旬中期以及8月下半月呈现阶段性增强; 在7月上旬末至中旬前期和7月下旬前中期出现阶段性减弱,分别与北方冷空气和台风活动影响相关联。对超前10 d以内的预报而言,大模型预报与实况总体一致,能预测出副热带高压对上海影响的4个增强阶段和2个减弱阶段,FourCastNet预测的副热带高压强度较其他模型稍弱。在次季节尺度上,随着超前预测时效的延长,大模型预测的副热带高压影响逐步减弱。尽管FuXi和Pangu超前11~15 d对盛夏影响上海上空的副热带高压演变仍有较好的预测效果(与实况的相关系数分别为0.69和0.62),但未能把握住7月下旬前中期的台风活动影响。在超前16~20 d时,除预测强度减弱外,Pangu预测的副热带高压对上海的影响变化不大,与实况的相关系数(0.26)通过了0.05信度的显著性检验,能反映出7月下半月—8月上旬以及8月下旬中后期的副热带高压影响。因此,相较高温预测而言,Pangu对2024年盛夏副热带高压影响的次季节尺度有效超前预测时效更长,可达16~20 d。
53个人工智能大模型超前不同时效对上海日最高气温的预测与实况(ERA5再分析资料)的对比(括号内数字表示相关系数,*、***分别表示通过0.05、0.001信度的显著性检验):(a)Pangu;(b)FourCastNet;(c)FuXi
Fig.5Comparison between forecasts of daily maximum temperature in Shanghai from three AI models at different lead times and observations (ERA5 reanalysis data) (numbers in parentheses indicate correlation coefficients, and * and *** represent coefficients statistically significant at the 0.05 and 0.001 levels, respectively) : (a) Pangu; (b) FourCastNet; (c) FuXi
进一步对各大模型不同超前时效的预测进行功率谱分析。对比实况(图9)发现,上述高温预测效果差异与不同模型对低频环流演变的预测效果不同相联系。上海上空500 hPa位势高度呈现10~20 d和30~60 d准周期(中心周期约为15 d和40 d)低频振荡,特别是10~20 d振荡更为显著(图9a)。对超前11~15 d的预测而言,FuXi和Pangu的预测仍存在中心周期分别为15 d和40 d的低频振荡(图9b、c),对超前11~15 d预测出副热带高压的阶段性演变有贡献。对Pangu而言,其超前16~20 d的预测仍能把握住上海上空位势高度呈现的准双周低频演变(图9d),这与其超前16~20 d能有效预测盛夏副热带高压影响一致。此外,由于超前16~20 d的预测对中心周期为40 d的位势高度低频振荡把握不足(图9d),导致其预测与实况的相关较超前11~15 d的预测明显减弱。
图9可见,观测和大模型中上海上空500 hPa位势高度呈现出显著的10~20 d准周期低频振荡。为进一步反映各模型对低频环流演变的把握能力,图10给出了各模型及传统的EC次季节模式超前11~15 d预测的10~20 d低频分量与观测的对比。可见,各模型、模式均能在超前11~15 d的时间尺度上给出有技巧(均通过0.05信度的显著性检验)的10~20 d低频分量预报。其中,FuXi和EC模式预测的上海上空局地环流的10~20 d低频分量与观测的相关最显著(通过0.005信度的显著性检验),其次是FourCastNet(通过0.02信度的显著性检验),再次是Pangu(通过0.05信度的显著性检验)。上述各模型、模式对环流低频分量的预测性能比较与其对实际环流的预测性能比较是一致的。例如,FuXi和EC模式超前11~15 d的预测与观测的相关系数分别达到0.69和0.68,再次是Pangu(0.62)和FourCastNet(0.2)。本文对上海Tmax低频分量的预测效果进行了类似的相关技巧检验(图略); 结果表明,FuXi超前11~15 d的预测与观测的相关系数为0.58(通过0.1信度的显著性检验),Pangu和EC模式的预测与观测的相关系数分别为0.54和0.52(通过0.15信度的显著性检验),FourCastNet的预测效果相对较差,相关系数仅为0.07。总体而言,模型对上海上空环流低频分量的超前预测效果越好,对Tmax的低频分量预测效果也越好。上述结果在一定程度上可解释FourCastNet超前11~15 d对实际Tmax的预测效果明显弱于其他模型的原因。综合图9图10可见,不同大模型对上海上空低频环流演变的反映能力越强,对盛夏影响上海的副热带高压的次季节预测效果越好,对其高温演变的次季节预测效果也越好。
6上海日最高气温实况(a.ERA5再分析资料)与3个大模型(b.Pangu; c.FourCastNet; d.FuXi)超前11~15 d预测结果的功率谱分析(实线表示功率谱,红、蓝虚线分别表示红噪声和95%置信水平)
Fig.6Power spectrum analysis of (a) observed daily maximum temperature (ERA5 reanalysis data) in Shanghai and (b—d) forecasts from three AI models with a lead time of 11—15 d: (b) Pangu, (c) FourCastNet, and (d) FuXi (the solid line represents the power spectrum; red and blue dashed lines indicate red noise and the95% confidence level, respectively)
73个人工智能大模型(a.Pangu; b.FourCastNet; c.FuXi)超前不同时效的日最高气温预测误差,以及上海日最高气温实况的演变(d; 蓝线为ERA5再分析资料,黄线为上海全市站点平均)
Fig.7Prediction errors of daily maximum temperature from three AI models: (a) Pangu, (b) FourCastNet, and (c) FuXi at different lead times, and (d) evolution of observed daily maximum temperature in Shanghai (blue line:ERA5 reanalysis data; yellow line:average from all basic observatories in Shanghai)
83个人工智能大模型(Pangu、FourCastNet、FuXi)超前不同时效预测的沿121°~122°E平均的500 hPa位势高度的演变(单位:gpm; a—c(e—g、h—j)分别为Pangu(FourCastNet、FuXi)超前1~5 d、6~10 d、11~15 d; d为Pangu超前16~20 d; k为实况)
Fig.8Evolution of 500 hPa geopotential height (units:gpm) along121°—122°E predicted by three AI models (Pangu, FourCastNet, and FuXi) at different lead times: (a—c, e—g, h—j) predictions from Pangu, FourCastNet, and FuXi, respectively, with lead times of 1—5 d, 6—10 d, and 11—15 d; (d) Pangu prediction with a lead time of 16—20 d; (k) observations
4 结论与讨论
针对人工智能大模型对超大城市极端高温次季节预测效果尚不清楚这一问题,本文以2024年上海盛夏为例,在分析该年盛夏高温背景及其低频演变的基础上,评估了3个人工智能气象大模型(Pangu、FuXi和FourCastNet)对上海高温及其环流的预测效果。主要结论如下:
1)2024年盛夏全国气温偏高,包括上海在内的长江流域尤为显著,与其上空的副热带高压强度呈显著正相关关系。上海的持续高温时段与10~20 d准周期振荡的正位相时段基本一致,并受到30~60 d低频振荡的调制,这与影响上海的副热带高压低频演变相联系。
2)不同大模型对2024年盛夏上海高温的次季节预测效果不同,部分模型(Pangu、FuXi)在提前15 d内可提供有技巧的预测参考。大模型对高温的次季节预测效果与其对高温低频演变的反映能力有关,对准双周和30~60 d低频振荡的预测越准确,对高温的次季节预测效果越好。该年出梅和台风影响阶段之外的大模型高温次季节预测效果更好,可为高温次季节预测提供更大的机会窗口。
9上海上空500 hPa位势高度实况(a.ERA5再分析资料)和人工智能大模型次季节预测结果(b.Pangu超前11~15 d预测; c.FuXi超前11~15 d预测; d.Pangu超前16~20 d预测)的功率谱分析(实线表示功率谱,红、蓝虚线分别表示红噪声和95%置信水平)
Fig.9Power spectrum analysis of (a) observed 500 hPa geopotential height (ERA5 reanalysis data) over Shanghai and (b—d) sub-seasonal predictions from AI models: (b) Pangu with a lead time of 11—15 d, (c) FuXi with a lead time of 11—15 d, and (d) Pangu with a lead time of 16—20 d (the solid line represents the power spectrum; red and blue dashed lines indicate red noise and the95% confidence level, respectively)
3)相较高温预测而言,部分大模型(Pangu)对2024年盛夏副热带高压影响的次季节尺度有效超前预测时效更长,可达16~20 d。大模型对上海上空低频环流演变的预测能力越强,对盛夏影响上海的副热带高压的次季节预测效果就越好。
本文的分析表明,不同的人工智能大模型对低频振荡的次季节预测能力仍有差异。以上海上空500 hPa位势高度低频演变为例,尽管3个大模型均能超前11~15 d给出有技巧的低频环流预测,但进一步计算发现各模型预测值的差异与初始状态有关(相关系数为-0.36)。换言之,各模型超前11~15 d对低频环流的预测差异在较大程度上受到初始低频状态的强度影响。当初始状态较弱时,各模型对低频环流的次季节预测差异更大,这种由初始状态带来的预测技巧的不确定性值得进一步关注。而且,不同大模型之间的结构差异对预测技巧的影响也值得进一步分析。同时,本文的人工智能大模型预测效果主要来自以2024年极端高温事件为例的评估,对其他极端高温事件的预测评估有待进一步开展,以期更全面地了解人工智能大模型对极端高温的预测性能,为改进大模型的次季节预测效果提供科学依据。此外,目前的大模型空间分辨率还不能满足城市精细化预测的需要。在改进大模型或动力模式对超大城市的高温次季节预测中,还需考虑城市化效应等下垫面的空间差异,以期提升预测的精细化程度。
10上海上空500 hPa位势高度(ERA5再分析资料)演变的10~20 d准周期振荡分量(LFO1)与人工智能大模型(Pangu、FuXi、FourCastNet)、EC次季节动力模式超前11~15 d的预测对比(图下方“Cor”表示预测和LFO1的相关系数,括号内数字表示显著性水平)
Fig.10Comparison of the10—20 d quasi-periodic oscillation component (LFO1) of 500 hPa geopotential height evolution (ERA5 reanalysis data) over Shanghai with predictions from AI large models (Pangu, FuXi, and FourCastNet) and the EC sub-seasonal dynamical model at a lead time of 11—15 d (“Cor” at the bottom of the figure indicates the correlation coefficient between predictions and LFO1, with numbers in parentheses indicating statistical significance levels)
12024年盛夏(7—8月)东亚气温距平的空间分布(a; 单位:℃)和1961—2024年上海全市盛夏平均气温的演变(b; 单位:℃)
Fig.1(a) Spatial distribution of temperature anomalies (units:℃) in East Asia during midsummer (July-August) 2024, and (b) evolution of the average temperature (units:℃) in Shanghai during midsummer from 1961 to 2024
22024年盛夏(7—8月)500 hPa位势高度异常(a; 阴影,单位:gpm; 细(粗)实线、细(粗)虚线分别表示5 760(5 880)gpm等值线及与之对应的气候平均值(1991—2020年平均)),以及去趋势(原始逐年序列减去其线性拟合项)的1961—2024年上海全市平均盛夏高温日数异常(红线,单位:d)及其上空(121°~122°E,30.5°~31.5°N)500 hPa位势高度异常(蓝线,单位:gpm)的演变(b)
Fig.2(a) Anomalies of the500 hPa geopotential height during midsummer (July-August) 2024 (shadings, units:gpm; thin (thick) solid and dashed lines represent the5 760 gpm (5 880 gpm) contours and the corresponding climatology from 1991 to 2020) , and (b) detrended evolution of anomalies in the average number of high-temperature days (red curve, units:d) in Shanghai and 500 hPa geopotential height (blue curve; units:gpm) over Shanghai (121°—122°E, 30.5°—31.5°N) during midsummer from 1961 to 2024
32024年盛夏(7—8月)上海全市平均日最高气温Tmax(a; 单位:℃)及其上空500 hPa位势高度z500(b; 单位:dagpm)的逐日演变及其低频分量(LFO1、LFO2分别表示10~20 d、30~60 d低频分量)
Fig.3Daily evolution and its low-frequency components of (a) average daily maximum temperature (units:℃) in Shanghai and (b) 500 hPa geopotential height (units:dagpm) over Shanghai during midsummer (July-August) 2024 (LFO1 and LFO2 represent low-frequency components with 10—20 d and 30—60 d periods, respectively)
4上海上空500 hPa位势高度的10~20 d低频分量与超前12 d(a)、9 d(b)、6 d(c)、3 d(d)和0 d(e)的10~20 d低频滤波后的500 hPa位势高度距平场的回归系数分布(单位:gpm; 交叉影线区域表示通过0.1信度的显著性检验);(f—j)同(a—e),但为Tmax低频分量与OLR的回归系数分布(单位:W/m2
Fig.4Spatial distribution of regression coefficients between the10—20 d low-frequency component of the500 hPa geopotential height over Shanghai and the anomaly fields of the10—20 d low-pass filtered 500 hPa geopotential height at lead times of (a) 12 d, (b) 9 d, (c) 6 d, (d) 3 d, and (e) 0 d (units:gpm; cross-hatched areas indicate statistically significant coefficients at the 0.1 level) . (f—j) as in (a—e) , but for the regression coefficients between the low-frequency component of Tmax and OLR (units:W/m2)
53个人工智能大模型超前不同时效对上海日最高气温的预测与实况(ERA5再分析资料)的对比(括号内数字表示相关系数,*、***分别表示通过0.05、0.001信度的显著性检验):(a)Pangu;(b)FourCastNet;(c)FuXi
Fig.5Comparison between forecasts of daily maximum temperature in Shanghai from three AI models at different lead times and observations (ERA5 reanalysis data) (numbers in parentheses indicate correlation coefficients, and * and *** represent coefficients statistically significant at the 0.05 and 0.001 levels, respectively) : (a) Pangu; (b) FourCastNet; (c) FuXi
6上海日最高气温实况(a.ERA5再分析资料)与3个大模型(b.Pangu; c.FourCastNet; d.FuXi)超前11~15 d预测结果的功率谱分析(实线表示功率谱,红、蓝虚线分别表示红噪声和95%置信水平)
Fig.6Power spectrum analysis of (a) observed daily maximum temperature (ERA5 reanalysis data) in Shanghai and (b—d) forecasts from three AI models with a lead time of 11—15 d: (b) Pangu, (c) FourCastNet, and (d) FuXi (the solid line represents the power spectrum; red and blue dashed lines indicate red noise and the95% confidence level, respectively)
73个人工智能大模型(a.Pangu; b.FourCastNet; c.FuXi)超前不同时效的日最高气温预测误差,以及上海日最高气温实况的演变(d; 蓝线为ERA5再分析资料,黄线为上海全市站点平均)
Fig.7Prediction errors of daily maximum temperature from three AI models: (a) Pangu, (b) FourCastNet, and (c) FuXi at different lead times, and (d) evolution of observed daily maximum temperature in Shanghai (blue line:ERA5 reanalysis data; yellow line:average from all basic observatories in Shanghai)
83个人工智能大模型(Pangu、FourCastNet、FuXi)超前不同时效预测的沿121°~122°E平均的500 hPa位势高度的演变(单位:gpm; a—c(e—g、h—j)分别为Pangu(FourCastNet、FuXi)超前1~5 d、6~10 d、11~15 d; d为Pangu超前16~20 d; k为实况)
Fig.8Evolution of 500 hPa geopotential height (units:gpm) along121°—122°E predicted by three AI models (Pangu, FourCastNet, and FuXi) at different lead times: (a—c, e—g, h—j) predictions from Pangu, FourCastNet, and FuXi, respectively, with lead times of 1—5 d, 6—10 d, and 11—15 d; (d) Pangu prediction with a lead time of 16—20 d; (k) observations
9上海上空500 hPa位势高度实况(a.ERA5再分析资料)和人工智能大模型次季节预测结果(b.Pangu超前11~15 d预测; c.FuXi超前11~15 d预测; d.Pangu超前16~20 d预测)的功率谱分析(实线表示功率谱,红、蓝虚线分别表示红噪声和95%置信水平)
Fig.9Power spectrum analysis of (a) observed 500 hPa geopotential height (ERA5 reanalysis data) over Shanghai and (b—d) sub-seasonal predictions from AI models: (b) Pangu with a lead time of 11—15 d, (c) FuXi with a lead time of 11—15 d, and (d) Pangu with a lead time of 16—20 d (the solid line represents the power spectrum; red and blue dashed lines indicate red noise and the95% confidence level, respectively)
10上海上空500 hPa位势高度(ERA5再分析资料)演变的10~20 d准周期振荡分量(LFO1)与人工智能大模型(Pangu、FuXi、FourCastNet)、EC次季节动力模式超前11~15 d的预测对比(图下方“Cor”表示预测和LFO1的相关系数,括号内数字表示显著性水平)
Fig.10Comparison of the10—20 d quasi-periodic oscillation component (LFO1) of 500 hPa geopotential height evolution (ERA5 reanalysis data) over Shanghai with predictions from AI large models (Pangu, FuXi, and FourCastNet) and the EC sub-seasonal dynamical model at a lead time of 11—15 d (“Cor” at the bottom of the figure indicates the correlation coefficient between predictions and LFO1, with numbers in parentheses indicating statistical significance levels)
Basu R P, 2009. High ambient temperature and mortality: a review of epidemiologic studies from 2001 to 2008[J]. Environ Health, 8: 40. DOI: 10. 1186/1476-069X-8-40.
Bi K F, Xie L X, Zhang H H, et al., 2023. Accurate medium-range global weather forecasting with 3D neural networks[J]. Nature, 619(7970): 533-538. DOI: 10. 1038/s41586-023-06185-3.
Charlton-Perez A J, Dacre H F, Driscoll S, et al., 2024. Do AI models produce better weather forecasts than physics-based models?A quantitative evaluation case study of Storm Ciarán[J]. npj Clim Atmos Sci, 7: 93. DOI: 10. 1038/s41612-024-00638-w.
Chen K, Han T, Gong J, et al., 2023. FengWu: pushing the skillful global medium-range weather forecast beyond 10 days lead[J]. arXiv: 2304. 02948. DOI: 10. 48550/arXiv. 2304. 02948.
Chen L, Zhong X H, Zhang F, et al., 2023. FuXi: a cascade machine learning forecasting system for 15-day global weather forecast[J]. npj Clim Atmos Sci, 6: 190. DOI: 10. 1038/s41612-023-00512-1.
Gasparrini A, Guo Y M, Hashizume M, et al., 2015. Mortality risk attributable to high and low ambient temperature: a multicountry observational study[J]. Lancet, 386(9991): 369-375. DOI: 10. 1016/S0140-6736(14)62114-0.
Goodfellow I, Bengio Y, Courville A, 2016. Deep learning[M]. Cambridge, Massachusetts: the MIT Press.
Grimmond C S B, Blackett M, Best M J, et al., 2010. The international urban energy balance models comparison project: first results from phase 1[J]. J Appl Meteor Climatol, 49(6): 1268-1292. DOI: 10. 1175/2010jamc2354. 1.
Hersbach H, Bell B, Berrisford P, et al., 2020. The ERA5 global reanalysis[J]. Quart J Roy Meteor Soc, 146(730): 1999-2049. DOI: 10. 1002/qj. 3803.
黄小猛, 林岩銮, 熊巍, 等, 2024. 数值预报AI气象大模型国际发展动态研究[J]. 大气科学学报, 47(1): 46-54. Huang X M, Lin Y L, Xiong W, et al., 2024. Research on international developments of AI large meteorological models in numerical forecasting[J]. Trans Atmos Sci, 47(1): 46-54. DOI: 10. 13878/j. cnki. dqkxxb. 20231201001. (in Chinese).
IPCC, 2021. Climate change 2021: the physical science basis[C]//Masson-Delmotte V, Zhai P, Pirani A, et al. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge: Cambridge University Press.
Lam R, Sanchez-Gonzalez A, Willson M, et al., 2023. Learning skillful medium-range global weather forecasting[J]. Science, 382(6677): 1416-1421. DOI: 10. 1126/science. adi2336.
雷蕾, 徐邦琪, 高庆九, 等, 2022. 基于卷积神经网络的长江流域夏季日最高温度延伸期预报方法研究[J]. 大气科学学报, 45(6): 835-849. Lei L, Hsu P C, Gao Q J, et al., 2022. Extended-range forecasting method of summer daily maximum temperature in the Yangtze River basin based on convolutional neural network[J]. Trans Atmos Sci, 45(6): 835-849. DOI: 10. 13878/j. cnki. dqkxxb. 20211101001. (in Chinese).
Liang P, Lin H, 2018. Sub-seasonal prediction over East Asia during boreal summer using the ECCC monthly forecasting system[J]. Climate Dyn, 50(3): 1007-1022. DOI: 10. 1007/s00382-017-3658-1.
Liang P, Lin H, Ding Y H, 2018. Dominant modes of sub-seasonal variability of East Asian summertime surface air temperature and their predictions[J]. J Climate, 31(7): 2729-2743. DOI: 10. 1175/jcli-d-17-0368. 1.
Liang P, Yan Z W, Li Z, 2022. Climatic warming in Shanghai during 1873—2019 based on homogenised temperature records[J]. Adv Clim Change Res, 13(4): 496-506. DOI: 10. 1016/j. accre. 2022. 05. 006.
Liang P, Zhang Z Q, Ding Y H, et al., 2024. The 2022 extreme heatwave in Shanghai, lower reaches of the Yangtze River valley: combined influences of multiscale variabilities[J]. Adv Atmos Sci, 41(4): 593-607. DOI: 10. 1007/s00376-023-3007-8.
Liang X, Jiang L P, Pan Y, et al., 2020. A 10-yr global land surface reanalysis interim dataset (CRA-Interim/Land): implementation and preliminary evaluation[J]. J Meteor Res, 34(1): 101-116. DOI: 10. 1007/s13351-020-9083-0.
Liebmann B, Smith C A, 1996. Description of a complete (interpolated) outgoing longwave radiation dataset[J]. Bull Amer Meteor Soc, 77: 1275-1277. DOI: 10. 1175/1520-0477-77. 6. 1274.
Liu D, Zhou R, Ma Q, et al., 2024. Spatio-temporal patterns and population exposure risks of urban heat island in megacity Shanghai, China[J]. Sustain Cities Soc, 108: 105500. DOI: 10. 1016/j. scs. 2024. 105500.
Ma Y, Liang P, Grimmond S, et al., 2022. Three-dimensional urban thermal effect across a large city cluster during an extreme heat wave: observational analysis[J]. J Meteor Res, 36(3): 387-400. DOI: 10. 1007/s13351-022-1171-x.
Molina M J, O'Brien T A, Anderson G, et al., 2023. A review of recent and emerging machine learning applications for climate variability and weather phenomena[J]. Artif Intell Earth Syst, 2(4): 1-17. DOI: 10. 1175/aies-d-22-0086. 1.
NASEM (National Academies of Sciences, Engineering, and Medicine), 2016. Next generation earth system prediction: strategies for subseasonal to seasonal forecasts[M]. Washington DC: the National Academies Press.
Oke T R, 1982. The energetic basis of the urban heat island[J]. Quart J Roy Meteor Soc, 108(455): 1-24. DOI: 10. 1002/qj. 49710845502.
Olivetti L, Messori G, 2024. Do data-driven models beat numerical models in forecasting weather extremes?A comparison of IFS HRES, Pangu-Weather, and GraphCast[J]. Geosci Model Dev, 17: 7915-7962. DOI: 10. 5194/gmd-17-7915-2024.
Pathak J, Subramanian S, Harrington P, et al., 2022. FourCastNet: a global data-driven high-resolution weather model using adaptive Fourier neural operators[J]. arXiv: 2202. 11214. DOI: 10. 48550/arXiv. 2202. 11214v1.
Rogers C D W, Kornhuber K, Perkins-Kirkpatrick S E, et al., 2022. Sixfold increase in historical Northern Hemisphere concurrent large heatwaves driven by warming and changing atmospheric circulations[J]. J Climate, 35(3): 1063-1078. DOI: 10. 1175/jcli-d-21-0200. 1.
Vitart F, Ardilouze C, Bonet A, et al., 2017. The sub-seasonal to seasonal (S2S) prediction project database[J]. Bull Amer Meteor Soc, 98(1): 163-173. DOI: 10. 1175/bams-d-16-0017. 1.
WMO/WWRP (World Meteorological Organization/World Weather Research Programme), 2024. Project Sub-seasonal Applications for aGriculture and Environment (SAGE) 2024—2028[EB/OL].(2024-09-18)[2025-01-08]. https://filecloud. wmo. int/share/s/QvoAiKfGSgqIVC1nM4W9sg.
Wu Z H, Huang N E, 2009. Ensemble empirical mode decomposition: a noise-assisted data analysis method[J]. Adv Adapt Data Anal, 1(1): 1-41. DOI: 10. 1142/s1793536909000047.
Xie J H, Hsu P C, Hu Y M, et al., 2024. Advancing subseasonal surface air temperature and heat wave prediction skill in China by incorporating scale interaction in a deep learning model[J]. Geophys Res Lett, 51(20): e2024GL111076. DOI: 10. 1029/2024gl111076.
Xu H X, Zhao Y, Zhao D J, et al., 2024. Improvement of disastrous extreme precipitation forecasting in North China by Pangu-weather AI-driven regional WRF model[J]. Environ Res Lett, 19(5): 054051. DOI: 10. 1088/1748-9326/ad41f0.
Yang J, Zhu T, Gao M N, et al., 2018. Late-July barrier for subseasonal forecast of summer daily maximum temperature over Yangtze River basin[J]. Geophys Res Lett, 45(22): 610-615. DOI: 10. 1029/2018gl080963.
杨秋明, 2018. 长江下游夏季低频温度和高温天气的延伸期预报研究[J]. 地球科学进展, 33(4): 385-395. Yang Q M, 2018. A study of the extended-range forecast for the low frequency temperature and high temperature weather over the lower reaches of Yangtze River valley in summer[J]. Advances in Earth Science, 33(4): 385-395. DOI: 10. 11867/j. issn. 1001-8166. 2018. 04. 0385. (in Chinese).
Zhu Z W, Li T, 2018. Extended-range forecasting of Chinese summer surface air temperature and heat waves[J]. Climate Dyn, 50(5): 2007-2021. DOI: 10. 1007/s00382-017-3733-7.

地址:江苏南京宁六路219号南京信息工程大学    邮编:210044

联系电话:025-58731158    E-mail:xbbjb@nuist.edu.cn    QQ交流群号:344646895

大气科学学报 ® 2026 版权所有  技术支持:北京勤云科技发展有限公司