Application of an Improved Analog-Based Heavy Precipitation Forecast Model to the Yangtze–Huai River Valley and Its Performance in June–July 2020

江淮流域强降水相似预报模型的改进及其在2020年6–7月强降水事件预报中的表现

+ Author Affiliations + Find other works by these authors
  • Corresponding author: Panmao ZHAI, pmzhai@cma.gov.cn
  • Funds:

    Supported by the National Natural Science Foundation of China (41905082), National Key Research and Development Program of China (2018YFC1507700), and Basic Research to Operation Fund of Chinese Academy of Meteorological Sciences (2019Y009)

  • doi: 10.1007/s13351-021-1059-1

PDF

  • Precipitation extremes, such as the record-breaking Meiyu characterized by frequent occurrences of rainstorms that resulted in severe flooding over the Yangtze–Huai River valley (YHRV) in June–July 2020, are always attracting considerable interest, highlighting the importance of improving the forecast accuracy at the medium-to-long range. In 2020, the Key Influential Systems based Analog Model (KISAM) developed in National Meteorological Center of China was further improved and brought into operational application, and its skill in forecasting heavy precipitation events (HPEs) of both long and short durations is analyzed in this study. Verification and comparison of this newly adapted analog model against the ECMWF ensemble mean forecasts at lead times of up to 15 days are carried out for the identified 16 HPEs over YHRV in June–July 2020. The results demonstrate that KISAM is advantageous over ECMWF ensemble mean for forecasts of heavy precipitation ≥ 25 mm day−1 at the medium-to-long (6–15-day) lead times, based on the traditional dichotomous metrics. However, at short lead times, ECMWF ensemble mean is advantageous over KISAM due largely to the low false alarm rates (FARs) benefited from an underestimation of the frequency of heavy precipitation. The analysis revealed that at the medium-to-long forecast range, the large fraction of misses induced by the high degree of under forecasting overwhelms the fairly good FARs in the ECMWF ensemble mean, which partly explains its inferiority to KISAM in terms of the threat score. Further assessment on forecasts of the latitudinal location of accumulated heavy precipitation indicates that smaller displacement errors also account for a part of the better performance of KISAM at lead times of 8–12 days.
    极端降水,因其强致灾性而备受关注并凸显中长期精准预报的重要性。为同时提高中国江淮流域持续时间长(≥ 3天)和短(1–2天)的强降水事件的预报技巧,国家气象中心基于关键影响系统的相似预报模型 (KISAM) 得到了改进并于2020年投入业务运行。本文检验对比了KISAM以及欧洲中期天气预报中心 (ECMWF) 在1–15天预报时效预报2020年6–7月江淮流域16个强降水事件的效果。结果表明,就传统二分类评分而言,KISAM 对强降水(25 mm day−1)在中长期预报时效(6–15天)的预报效果优于 ECMWF;但在短期预报时效,ECMWF集合平均预报优于KISAM,因为ECMWF对强降水频率低估导致较低的错报率。在中长期预报时效,ECMWF 集合平均预报对强降水频率的显著低估在一定程度上解释了其TS 评分相比 KISAM的劣势。对刻画强降水事件雨带纬度位置能力的评估表明,较小的雨带位置误差也是 KISAM 在 8–12 天预报时效评分更佳的部分原因。
  • 加载中
  • Fig. 1.  Schematic flow diagram for the training and real-time forecasting processes in the analog forecast model (KISAM). The light blue and brown arrows indicate the training and real-time forecasting processes, respectively(请作者核查并确认该图题调整得是否妥当).

    Fig. 2.  The difference of average TSs between the KISAM and ECMWF ensemble mean in the hindcasts of the 44 HPEs over 2016–2018 at the lead times of up to 15 days(请作者再次确认纵坐标名称及是否有单位).

    Fig. 3.  (a) The time–latitude cross section of daily precipitation (mm day−1) averaged along 112°–121°E from 1 June to 31 July and (b) proportion of stations with daily precipitation exceeding 50 mm in YHRV (26°–34°N, 112°–121°E). The horizontal red lines and numbers in (b) indicate the 16 HPEs identified by the object-oriented method(请作者补充蓝色和橙色图注在图b中;且the black dashed line代表什么?).

    Fig. 4.  Average TSs of (a) KISAM and ECMWF ensemble mean, and (c) frequencies of when KISAM and ECMWF ensemble mean score higher than each other or both score zero for the forecasting of daily precipitation ≥ 25 mm day−1 during the 16 HPEs at 1–15 lead days. (b) and (d) are the same as (a) and (c), but for KISAM_ms. The numbers of HPEs used in the verification on each lead day are indicated on the top axis(请作者再次确认大图题及各个小图的纵坐标名称及是否有单位).

    Fig. 5.  Categorical performance diagram for KISAM and ECMWF ensemble mean in forecasting precipitation no less than 25 mm day−1 at the lead times of 1–15 days during the 16 HPEs. The horizontal and vertical axes indicate SR and POD, respectively. Blue dashed lines represent BIAS with labels on the outward extension of the line, while labeled yellow solid contours are TSs(请作者确认图中重叠数字,是否妥当).

    Fig. 6.  Latitudinal location of the observed zonal-mean accumulated precipitation (mm) averaged over 112°–121°E and the corresponding forecasts by (a, b) ECMWF, (c, d) KISAM and (e, f) KISAM_ms for the two HPEs during (left column) 21–25 June at 1–11 lead days and (right column) 4–9 July at 1–10 lead days.

    Fig. 7.  Mean absolute error in the forecasts from ECMWF ensemble mean and KISAM at 1–15 lead days for the latitudinal location of the heaviest center of the zonal-mean precipitation accumulation along 112°–121°E during the 16 HPEs. The numbers of HPEs used in the verification on each lead day are indicated on the top axis.

    Table 1.  Parameters derived from the training process(1. 疑问同图1,P S 均斜体?ps: 变量需用斜体。qu,qv,cv统一用大写还是小写?200,500等用下标么?2. 该表格缺表头,各个P的表头是Parameter,那么第二行对应的数字用什么表头?)

    P200P500P700quP700qvScv
    0.14160.16240.44670.25930.6231
    Download: Download as CSV

    Table 2.  Contingency table

    Observation
    ForecastYesNo
    YesHitFalse alarm
    NoMissCorrect negative
    Download: Download as CSV
  • [1]

    Ben Daoud, A., E. Sauquet, G. Bontron, et al., 2016: Daily quantitative precipitation forecasts based on the analogue method: Improvements and application to a French large river basin. Atmos. Res., 169, 147–159. doi: 10.1016/j.atmosres.2015.09.015.
    [2]

    Bougeault, P., Z. Toth, C. Bishop, et al., 2010: The THORPEX Interactive Grand Global Ensemble. Bull. Amer. Meteor. Soc., 91, 1059–1072. doi: 10.1175/2010bams2853.1.
    [3]

    Chen, T., F. H. Zhang, C. Yu, et al., 2020: Synoptic analysis of extreme Meiyu precipitation over Yangtze River Basin during June–July 2020. Meteor. Mon., 46, 1415–1426. doi: 10.7519/j.issn.1000-0526.2020.11.003. (in Chinese)
    [4]

    Chen, Y., and P. M. Zhai, 2013: Persistent extreme precipitation events in China during 1951–2010. Clim. Res., 57, 143–155. doi: 10.3354/cr01171.
    [5]

    Ding, Y. H., Z. Y. Wang, and Y. Sun, 2008: Inter-decadal variation of the summer precipitation in East China and its association with decreasing Asian summer monsoon. Part I: Observed evidences. Int. J. Climatol., 28, 1139–1161. doi: 10.1002/joc.1615.
    [6]

    Duan, W. L., B. He, D. Nover, et al., 2016: Floods and associated socioeconomic damages in China over the last century. Nat. Hazards, 82, 401–413. doi: 10.1007/s11069-016-2207-2.
    [7]

    Gibergans-Báguena, J., and M. C. Llasat, 2007: Improvement of the analog forecasting method by using local thermodynamic data. Application to autumn precipitation in Catalonia. Atmos. Res., 86, 173–193. doi: 10.1016/j.atmosres.2007.04.002.
    [8]

    Huang, L., and Y. L. Luo, 2017: Evaluation of quantitative precipitation forecasts by TIGGE ensembles for south China during the presummer rainy season. J. Geophys. Res. Atmos., 122, 8494–8516. doi: 10.1002/2017jd026512.
    [9]

    Kalnay, E., M. Kanamitsu, R. Kistler, et al., 1996: The NCEP/NCAR 40-year reanalysis project. Bull. Amer. Meteor. Soc., 77, 437–471. doi: 10.1175/1520-0477(1996)077<0437:tnyrp>2.0.co;2.
    [10]

    Krishnamurti, T. N., A. D. Sagadevan, A. Chakraborty, et al., 2009: Improving multimodel weather forecast of monsoon rain over China using FSU superensemble. Adv. Atmos. Sci., 26, 813–839. doi: 10.1007/s00376-009-8162-z.
    [11]

    Liu, B. Q., Y. H. Yan, C. W. Zhu, et al., 2020: Record-breaking Meiyu rainfall around the Yangtze River in 2020 regulated by the subseasonal phase transition of the North Atlantic Oscillation. Geophys. Res. Lett., 47, e2020GL090342. doi: 10.1029/2020gl090342.
    [12]

    Liu, L., C. Gao, Q. Zhu, et al., 2019: Evaluation of TIGGE daily accumulated precipitation forecasts over the Qu River basin, China. J. Meteor. Res., 33, 747–764. doi: 10.1007/s13351-019-8096-z.
    [13]

    Liu, Y. Y., and Y. H. Ding, 2020: Characteristics and possible causes for the extreme Meiyu in 2020. Meteor. Mon., 46, 1393–1404. doi: 10.7519/j.issn.1000-0526.2020.11.001. (in Chinese)
    [14]

    Lorenz, E. N., 1969: Atmospheric predictability as revealed by naturally occurring analogues. J. Atmos. Sci., 26, 636–646. doi: 10.1175/1520-0469(1969)26<636:aparbn>2.0.co;2.
    [15]

    Luo, Y. L., J. S. Sun, Y. Li, et al., 2020: Science and prediction of heavy rainfall over China: Research progress since the reform and opening-up of new China. J. Meteor. Res., 34, 427–459. doi: 10.1007/s13351-020-0006-x.
    [16]

    Matulla, C., X. Zhang, X. L. Wang, et al., 2008: Influence of similarity measures on the performance of the analog method for downscaling daily precipitation. Climate Dyn., 30, 133–144. doi: 10.1007/s00382-007-0277-2.
    [17]

    Mehrotra, R., J. P. Evans, A. Sharma, et al., 2014: Evaluation of downscaled daily rainfall hindcasts over Sydney, Australia using statistical and dynamical downscaling approaches. Hydrol. Res., 45, 226–249. doi: 10.2166/nh.2013.094.
    [18]

    Niu, R. Y., C. H. Liu, W. Y. Liu, et al., 2018: Characteristics of temporal and spatial distribution of regional rainstorm processes to the East of 95°E in China during 1981–2015. Acta Meteor. Sinica, 76, 182–195. (in Chinese)
    [19]

    Ralph, F. M., E. Sukovich, D. Reynolds, et al., 2010: Assessment of extreme quantitative precipitation forecasts and development of regional extreme event thresholds using data from HMT-2006 and COOP observers. J. Hydrometeorol., 11, 1286–1304. doi: 10.1175/2010jhm1232.1.
    [20]

    Ran, Q. H., W. Fu, Y. Liu, et al., 2018: Evaluation of quantitative precipitation predictions by ECMWF, CMA, and UKMO for flood forecasting: Application to two basins in China. Nat. Hazards Rev., 19, 05018003. doi: 10.1061/(asce)nh.1527-6996.0000282.
    [21]

    Roebber, P. J., 2009: Visualizing multiple measures of forecast quality. Wea. Forecasting, 24, 601–608. doi: 10.1175/2008waf2222159.1.
    [22]

    Scheuerer, M., and T. M. Hamill, 2015: Statistical postprocessing of ensemble precipitation forecasts by fitting censored, shifted gamma distributions. Mon. Wea. Rev., 143, 4578–4596. doi: 10.1175/mwr-d-15-0061.1.
    [23]

    Sharma, S., R. Siddique, N. Balderas, et al., 2017: Eastern U.S. verification of ensemble precipitation forecasts. Wea. Forecasting, 32, 117–139. doi: 10.1175/waf-d-16-0094.1.
    [24]

    Su, X., H. L. Yuan, Y. J. Zhu, et al., 2014: Evaluation of TIGGE ensemble predictions of Northern Hemisphere summer precipitation during 2008–2012. J. Geophys. Res. Atmos., 119, 7292–7310. doi: 10.1002/2014jd021733.
    [25]

    Sukovich, E. M., F. M. Ralph, F. E. Barthold, et al., 2014: Extreme quantitative precipitation forecast performance at the Weather Prediction Center from 2001 to 2011. Wea. Forecasting, 29, 894–911. doi: 10.1175/waf-d-13-00061.1.
    [26]

    Xie, Z. Q., Y. Du, Y. Zeng, et al., 2018: Classification of yearly extreme precipitation events and associated flood risk in the Yangtze-Huaihe River Valley. Sci. China Earth Sci., 61, 1341–1356. doi: 10.1007/s11430-017-9212-8.
    [27]

    Yang, X.-S., and S. Deb, 2009: Cuckoo search via Lévy flights. Proceedings of 2009 World Congress on Nature & Biologically Inspired Computing (NaBIC), Coimbatore, India, 9–11 December, IEEE, Coimbatore, India, 210–214, doi: 10.1109/NABIC.2009.5393690.
    [28]

    Yang, X.-S., and S. Deb, 2014: Cuckoo search: recent advances and applications. Neural Comput. Appl., 24, 169–174. doi: 10.1007/s00521-013-1367-1.
    [29]

    Zhai, P. M., Y. Q. Ni, and Y. Chen, 2013: Mechanism and forecasting method of persistent extreme weather events: Review and prospect. Adv. Earth Sci., 28, 1177–1188. (in Chinese)
    [30]

    Zhang, F. H., T. Chen, F. Zhang, et al., 2020: Extreme features of severe precipitation in Meiyu period over the middle and lower reaches of Yangtze River Basin in June–July 2020. Meteor. Mon., 46, 1405–1414. doi: 10.7519/j.issn.1000-0526.2020.11.002. (in Chinese)
    [31]

    Zhang, R. H., 2015: Changes in East Asian summer monsoon and summer rainfall over eastern China during recent decades. Sci. Bull., 60, 1222–1224. doi: 10.1007/s11434-015-0824-x.
    [32]

    Zhou, B. Q., and P. M. Zhai, 2016: A new forecast model based on the analog method for persistent extreme precipitation. Wea. Forecasting, 31, 1325–1341. doi: 10.1175/waf-d-15-0174.1.
    [33]

    Zhou, B. Q., P. M. Zhai, and R. Y. Niu, 2018: Comparative assessment of two objective forecast models for cases of persistent extreme precipitation events in the Yangtze–Huai River valley in summer 2016. Wea. Forecasting, 33, 221–238. doi: 10.1175/waf-d-17-0039.1.
    [34]

    Zhou, T. J., D. Y. Gong, J. Li, et al., 2009: Detecting and understanding the multi-decadal variability of the East Asian Summer Monsoon Recent progress and state of affairs. Meteor. Z., 18, 455–467. doi: 10.1127/0941-2948/2009/0396.
    [35]

    Zorita, E., and H. Von Storch, 1999: The analog method as a simple statistical downscaling technique: Comparison with more complicated methods. J. Climate, 12, 2474–2489. doi: 10.1175/1520-0442(1999)012<2474:tamaas>2.0.co;2.
  • 加载中
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Application of an Improved Analog-Based Heavy Precipitation Forecast Model to the Yangtze–Huai River Valley and Its Performance in June–July 2020

    Corresponding author: Panmao ZHAI, pmzhai@cma.gov.cn
  • 1. State Key Laboratory of Severe Weather, Chinese Academy of Meteorological Sciences, China Meteorological Administration, Beijing 100081
  • 2. National Meteorological Center, China Meteorological Administration, Beijing 100081
Funds: Supported by the National Natural Science Foundation of China (41905082), National Key Research and Development Program of China (2018YFC1507700), and Basic Research to Operation Fund of Chinese Academy of Meteorological Sciences (2019Y009)

Abstract: Precipitation extremes, such as the record-breaking Meiyu characterized by frequent occurrences of rainstorms that resulted in severe flooding over the Yangtze–Huai River valley (YHRV) in June–July 2020, are always attracting considerable interest, highlighting the importance of improving the forecast accuracy at the medium-to-long range. In 2020, the Key Influential Systems based Analog Model (KISAM) developed in National Meteorological Center of China was further improved and brought into operational application, and its skill in forecasting heavy precipitation events (HPEs) of both long and short durations is analyzed in this study. Verification and comparison of this newly adapted analog model against the ECMWF ensemble mean forecasts at lead times of up to 15 days are carried out for the identified 16 HPEs over YHRV in June–July 2020. The results demonstrate that KISAM is advantageous over ECMWF ensemble mean for forecasts of heavy precipitation ≥ 25 mm day−1 at the medium-to-long (6–15-day) lead times, based on the traditional dichotomous metrics. However, at short lead times, ECMWF ensemble mean is advantageous over KISAM due largely to the low false alarm rates (FARs) benefited from an underestimation of the frequency of heavy precipitation. The analysis revealed that at the medium-to-long forecast range, the large fraction of misses induced by the high degree of under forecasting overwhelms the fairly good FARs in the ECMWF ensemble mean, which partly explains its inferiority to KISAM in terms of the threat score. Further assessment on forecasts of the latitudinal location of accumulated heavy precipitation indicates that smaller displacement errors also account for a part of the better performance of KISAM at lead times of 8–12 days.

江淮流域强降水相似预报模型的改进及其在2020年6–7月强降水事件预报中的表现

极端降水,因其强致灾性而备受关注并凸显中长期精准预报的重要性。为同时提高中国江淮流域持续时间长(≥ 3天)和短(1–2天)的强降水事件的预报技巧,国家气象中心基于关键影响系统的相似预报模型 (KISAM) 得到了改进并于2020年投入业务运行。本文检验对比了KISAM以及欧洲中期天气预报中心 (ECMWF) 在1–15天预报时效预报2020年6–7月江淮流域16个强降水事件的效果。结果表明,就传统二分类评分而言,KISAM 对强降水(25 mm day−1)在中长期预报时效(6–15天)的预报效果优于 ECMWF;但在短期预报时效,ECMWF集合平均预报优于KISAM,因为ECMWF对强降水频率低估导致较低的错报率。在中长期预报时效,ECMWF 集合平均预报对强降水频率的显著低估在一定程度上解释了其TS 评分相比 KISAM的劣势。对刻画强降水事件雨带纬度位置能力的评估表明,较小的雨带位置误差也是 KISAM 在 8–12 天预报时效评分更佳的部分原因。
1.   Introduction
  • Along with weakening of the East Asian summer monsoon (EASM), the major monsoonal rainband shifts southward in the late 1970s, which results in more precipitation over the Yangtze–Huai River valley (YHRV; Ding et al., 2008; Zhou et al., 2009; Zhang, 2015). Consequently, in the past several decades, YHRV has suffered great loss in life and property from increasingly frequent floods, accounting for up to 39.9% of total floods in China (Duan et al., 2016; Xie et al., 2018; Zhang et al., 2020). Unfortunately, June–July 2020 had witnessed another record-breaking Meiyu season since 1961, in terms of both persistence and accumulated amount of precipitation (Liu et al., 2020; Liu and Ding, 2020). The consecutive occurrence of heavy rainfall events led to severe flooding in YHRV in 2020 and aroused great concern from the government, public, and research community. For better decision-making that helps prevent damage to this densely populated and agricultural region, extending the valid forecast lead time for heavy precipitation to the medium-to-long range is of great significance (Bougeault et al., 2010; Zhai et al., 2013). Although the forecast accuracy of numerical weather prediction (NWP) models has been steadily improving, the skill in quantitative precipitation forecasts at longer lead times is still acknowledged to be insufficient (Sukovich et al., 2014; Luo et al., 2020). The large uncertainty in forecast of heavy precipitation poses a grand challenge for operational forecasters (Scheuerer and Hamill, 2015).

    The general underestimation of heavy precipitation for the forecasts from ensemble prediction systems (EPSs) collected in the Observing System Research and Predictability Experiment (THORPEX) Interactive Grand Global Ensemble (TIGGE) has been reported in a number of verification studies (Ralph et al., 2010; Su et al., 2014; Sharma et al., 2017). Quantitative precipitation forecasts for various regions and basins in China illustrate no exception. Huang and Luo (2017) revealed an underestimation of heavy rainfall [> 30 mm (12 h)−1] for forecasts generated by EPSs from several operational weather forecasting centers in South China during the presummer rainy season. Though the monsoon rainfall forecast is improved by a superensemble strategy, heavy precipitation during the onset of the South China Sea monsoon and Meiyu is still clearly underestimated in the superensemble forecasts (Krishnamurti et al., 2009). For the Linxian basin in northern China and the Qu River basin in eastern China, heavy rainfall defined by various thresholds is also underforecasted by EPSs from TIGGE (Ran et al., 2018; Liu et al., 2019). For the forecasting of two persistent extreme precipitation events over YHRV in summer 2016, the generally more skillful EPS from the ECMWF not only under-predicted heavy precipitation but also placed the rainband more northward than observation at longer lead days (Zhou et al., 2018). Therefore, improving heavy rainfall forecast is still a high priority for operational forecast services and meteorological research (Sukovich et al., 2014).

    As a common way to alleviate the deficiency of low-resolution global NWP models in forecasting heavy precipitation, statistical downscaling is cheap and easy to implement, such as the analog method (Lorenz, 1969; Zorita and Von Storch, 1999; Ben Daoud et al., 2016). The Key Influential Systems based Analog Model (KISAM) formerly developed by Zhou and Zhai (2016) has shown an advantage over the direct model output from ECMWF for the forecasting of heavy precipitation at lead times longer than three days over YHRV. However, its applicability is limited to persistent extreme precipitation, which is defined as 50 mm or larger daily precipitation at no less than 3 neighboring stations lasting at least 3 days (Chen and Zhai, 2013). Due to the fact that precipitation events meeting this rigorous criterion over YHRV are rare, KISAM is used infrequently in operational applications. Furthermore, since KISAM is optimized by using the persistent extreme precipitation cases, poor performance is found in the forecasting of heavy precipitation cases with duration shorter than three days. Motivated by this, refining KISAM to extend its applicability to the forecasting of heavy precipitation events (HPEs) over YHRV with both short and prolonged durations was carried out. To include the HPEs with short duration in the historical pool, the definition for the HPE from Niu et al. (2018) was adopted (details are described in the following section), which relaxes the criterion of a minimum 3-day duration in Chen and Zhai (2013) and only requires heavy precipitation occurring at no less than 15 neighboring grids over YHRV. The enriched historical archive of HPEs was then used to rebuild KISAM. This paper aims to introduce the adaptation of KISAM and verify the performance of the improved KISAM in forecasting the HPEs over YHRV in June–July 2020, with a comparison to the ensemble mean precipitation forecast from ECMWF.

    The remainder of the paper is organized as follows. Data employed in this study and the refinement of KISAM, as well as the verification methods, are described in Section 2. Brief introductions of the HPEs in June–July 2020 are presented in Section 3, followed by the verification results in Section 4. This paper ends with conclusions and outlook in Section 5.

2.   Data and methods
  • The daily reanalysis data at 2.5° × 2.5° horizontal resolution, obtained from the NCEP–NCAR (Kalnay et al., 1996), are used to characterize the large-scale atmospheric circulation on multiple levels in the training process of an analog model. For real-time operational forecasts in 2020, the forecasted large-scale information by ECMWF at 1200 UTC for the next 1–15 days is used as input for the analog model. The variables include zonal wind at 200 hPa; geopotential height at 500 hPa; and specific humidity as well as zonal and meridional winds at 700 hPa. In the training and validation processes, daily precipitation observations from 1200 to 1200 UTC of the following day at 430 stations in YHRV (26°–34°N, 112°–121°E) during 1981–2020 are obtained from the National Meteorological Information Center, China Meteorological Administration (CMA). The ensemble mean 1–15-day precipitation forecasts from ECMWF initialized at 1200 UTC each day are used for comparison, which are derived by equally weighting all the 51 members. To facilitate the comparison, the observed precipitation data during 1981–2020 are interpolated onto a 1° × 1° grid, in common with that of the ensemble mean forecasts from ECMWF.

  • Following the methodological flow in Zhou and Zhai (2016), KISAM is rebuilt with an enriched historical archive of HPEs over YHRV and circulation sequence with the shortened length. As illustrated in the schematic flow chart (Fig. 1), the zonal wind at 200 hPa in the area of 25°–55°N and 70°–160°E, the geopotential height at 500 hPa covering 0–70°N and 30°–180°E, as well as the zonal and meridional water vapor transport at 700 hPa over 0–35°N and 70°–160°E, are selected to characterize the large-scale circulation. Since precipitation at a given day is considered to be also influenced by circulation at the proceeding days, circulation sequences including large-scale information from the target day and bygone days are used as the predictors (Matulla et al., 2008). The length of the circulation sequence is shortened to 3 days, instead of the 8 days used in Zhou and Zhai (2016), considering that the HPEs in the historical pool will include those with short duration which only lasts 1 or 2 days. For a target circulation sequence, analogs are searched from the 730 candidates of the 3-day circulation sequence in the historical pool, corresponding to the 730 total days belonging to HPEs over the warm seasons (May–September) of 1981–2015 in YHRV identified through the object-oriented method in Niu et al. (2018). This method firstly interpolates precipitation data to a 0.5° × 0.5° grid and then identifies a heavy precipitation day when there is 50 mm day−1 or more precipitation occurring at no less than 15 neighboring grids over YHRV. Taking each heavy precipitation day as an object, subjective diagnoses are carried out on the synoptic processes in the adjacent 3–7 days ahead or behind, through daily synoptic charts comprising geopotential heights, wind, and relative humidity at multiple levels and surface pressure, along with the daily precipitation amount and identified heavy precipitation on the grid. The onset, development, weakening, and retreat of the synoptic processes governing heavy precipitation are subjectively analyzed and determined according to the principles of synoptic meteorology. Heavy precipitation days belonging to the same synoptic process are grouped into one HPE. Besides the HPEs during 1981–2015 utilized for training, 44 cases over 2016–2018 are preserved for independent validation. The analog metric used is the weight-assigned Cosine similarity, which is calculated as:

    Figure 1.  Schematic flow diagram for the training and real-time forecasting processes in the analog forecast model (KISAM). The light blue and brown arrows indicate the training and real-time forecasting processes, respectively(请作者核查并确认该图题调整得是否妥当).

    $${\rm{Similarity}} = \frac{{A*B}}{{\left\| A \right\|*\left\| B \right\|}} = \frac{{\sum\limits_{i = 1}^n {G(i){A_i}} G(i){B_i}}}{{\sqrt {\sum\limits_{i = 1}^n {{{[G(i){A_i}]}^2}} } \sqrt {\sum\limits_{i = 1}^n {{{[G(i){B_i}]}^2}} } }},$$ (1)

    where A and B represent two circulation fields, and n is the total number of grid points, G(i) is the weight function which is defined as the normalized absolute correlation coefficient between each predictor at each grid point and the regional mean precipitation over YHRV during the training period. The similarity between the circulation sequences is derived from the weighted sum of the similarities from the target day and bygone two days(请作者确认上述2处between是否需要改用among).

    An integrated similarity for the circulation sequences is then combined from the four similarities of the four predictors with normalized weights assigned (P1, P2, P3, P4, shown in the green box in Fig. 1). The three most similar analogs are selected to get an ensemble mean heavy precipitation reforecast, if the integrated similarities exceed the critical threshold (Scv). Using leave-one-out cross validation and an optimization method called Cuckoo Search (Yang and Deb, 2009, 2014), KISAM was retrained and optimized. In the training process, the normalized weights assigned to the four predictors and the critical threshold judging whether any historical record could be identified as an analog were determined, as shown in Table 1. In the independent validation period, the skill of KISAM for forecasting 25 mm day−1 or more precipitation is lower than ECMWF ensemble mean at 1–4 lead days in terms of the threat score (TS). Similar to what has been discussed in Zhou and Zhai (2016), the deficiency of KISAM at the first four lead days is possibly resulted from the predictors used in KISAM are all large-scale variables(请作者核查语言语法表达). KISAM does not take variables related to meso-scale and small-scale systems into consideration, which matters for the short-term precipitation forecast. However, KISAM shows an advantage over the ECMWF ensemble mean for forecasting heavy precipitation at lead times longer than four days and is thus demonstrated to be a skillful analog model at the medium-to-long range (Fig. 2). Thus, the new KISAM aimed at forecasting HPEs with whether the short or long duration over YHRV was established. With the input of 1–15-day circulation forecasts from ECMWF, KISAM will issue real-time heavy precipitation forecasts for the next 15 days if 3 or more analogs could be matched (Fig. 1). Two kinds of forecast products are output in real-time operations—One is the precipitation from the weighted ensemble mean of the three most similar analogs (KISAM) and the one is the precipitation from the most similar analog (KISAM_ms). KISAM_ms is used to provide information on the precipitation intensity, considering the possible smoothing effect of the ensemble mean.

    P200P500P700quP700qvScv
    0.14160.16240.44670.25930.6231

    Table 1.  Parameters derived from the training process(1. 疑问同图1,P S 均斜体?ps: 变量需用斜体。qu,qv,cv统一用大写还是小写?200,500等用下标么?2. 该表格缺表头,各个P的表头是Parameter,那么第二行对应的数字用什么表头?)

    Figure 2.  The difference of average TSs between the KISAM and ECMWF ensemble mean in the hindcasts of the 44 HPEs over 2016–2018 at the lead times of up to 15 days(请作者再次确认纵坐标名称及是否有单位).

  • Scores calculated from the contingency table are used to verify the performance of KISAM and ECMWF in the heavy precipitation forecast, including the success ratio (SR), probability of detection (POD), threat score (TS), and frequency bias (BIAS). They are calculated as:

    $$ {\rm{SR}}=\frac{\mathrm{h}\mathrm{i}\mathrm{t}}{\mathrm{h}\mathrm{i}\mathrm{t}+\mathrm{f}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}\;\mathrm{a}\mathrm{l}\mathrm{a}\mathrm{r}\mathrm{m}}, $$ (2)
    $$ {\rm{POD}}=\frac{\mathrm{h}\mathrm{i}\mathrm{t}}{\mathrm{h}\mathrm{i}\mathrm{t}+\mathrm{m}\mathrm{i}\mathrm{s}\mathrm{s}}, $$ (3)
    $$ {\rm{TS}}=\frac{\mathrm{hit}}{\mathrm{hit}+\mathrm{false}\;\mathrm{alarm}+\mathrm{miss}}=\frac{1}{\frac{1}{\rm POD}+\frac{1}{\rm SR}-1}, $$ (4)
    $$ {\rm{BIAS}}=\frac{\mathrm{h}\mathrm{i}\mathrm{t}+\mathrm{f}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}\;\mathrm{a}\mathrm{l}\mathrm{a}\mathrm{r}\mathrm{m}}{\mathrm{h}\mathrm{i}\mathrm{t}+\mathrm{m}\mathrm{i}\mathrm{s}\mathrm{s}}=\frac{\rm POD}{\rm SR}, $$ (5)

    where the hit, miss, false alarm, and correct negative come from the contingency table (Table 2). Taking precipitation at a given threshold as an event, SR and POD measure the fraction that forecasted “yes” events are correctly observed and the observed events are well forecasted respectively. TS indicates the ratio that observed or forecasted “yes” events are correctly forecasted(请作者核查上述两处语言,是定语从句么?若是,不能用that做关系词). BIAS indicates whether the forecast tends to yield more (BIAS > 1) or less (BIAS < 1) events. The better the forecast, the closer to unity the four metrics. Since TS and BIAS could be calculated from SR and POD, they can be synthesized and illustrated in a single diagram for ease of interpreting the statistics (Roebber, 2009). The larger SR and POD mean the smaller fraction of false alarms and misses, respectively.

    Observation
    ForecastYesNo
    YesHitFalse alarm
    NoMissCorrect negative

    Table 2.  Contingency table

    In addition, the meridional location of rainband is verified by visually comparing the observed and forecasted zonal-mean accumulated precipitation over 112°–121°E during an HPE. To quantify the displacement of the forecasted rainband, the mean absolute error for forecasting the latitudinal location of the heaviest center of the zonal-mean precipitation accumulation is employed. It needs to be noted that the forecast lead time for an HPE in this paper refers to the number of days in advance of the first day of the event. Therefore, for forecasts with lead times as long as 15 days, the longest lead time for predicting an x-day HPE could only be up to 16 minus x days.

3.   HPEs in June–July 2020
  • Precipitation in June–July 2020 over YHRV was characterized by the long persistence, wide extent, and record-breaking accumulated amount (Chen et al., 2020; Zhang et al., 2020). From the time–latitude cross section of the daily precipitation along 112°–121°E, it can be seen that, although the rainband showed north–south oscillation, heavy precipitation fell steadily in YHRV (Fig. 3a). Out of the 61 days in June and July, 30 days underwent precipitation no less than 50 mm day−1 over at least 5% stations in YHRV (Fig. 3b). According to Niu et al. (2018), days with 50 mm or more precipitation accumulation at no less than 15 neighboring grids (0.5° × 0.5° grid) over YHRV are first identified objectively. Based on subjectively examining the onset, evolution, and retreat of key synoptic processes from the daily synoptic charts, the identified heavy precipitation days belonging to the same synoptic process are classified into one HPE. Ultimately, a total of 16 HPEs are identified over YHRV in June–July 2020, durations of which are marked by the red horizontal lines in Fig. 3b. Some of these HPEs are connected with no interval but are not grouped into one event because the key influential synoptic process has altered. At the most concentrated stage of precipitation from 10 June to 22 July, the longest interval between the HPEs is only 3 days (Fig. 3b). Based on the 16 HPEs over YHRV in June–July 2020, the skills of the refined KISAM and the NWP model from ECMWF in forecasting heavy precipitation at the medium-to-long range are evaluated.

    Figure 3.  (a) The time–latitude cross section of daily precipitation (mm day−1) averaged along 112°–121°E from 1 June to 31 July and (b) proportion of stations with daily precipitation exceeding 50 mm in YHRV (26°–34°N, 112°–121°E). The horizontal red lines and numbers in (b) indicate the 16 HPEs identified by the object-oriented method(请作者补充蓝色和橙色图注在图b中;且the black dashed line代表什么?).

4.   Performance of KISAM and ECMWF
  • As what has been pointed out in the section on verification methods, for the 15-day forecasts, an x-day HPE could be forecasted as long as 16 minus x days in advance. The duration of the 16 HPEs identified in YHRV ranges from 1 to 6 days, indicating that the longest forecast lead time of these HPEs varies from 10 to 15 days. Therefore, forecast performance could be validated for all the 16 HPEs at lead times of up to 10 days. At lead times longer than 10 days, the number of HPEs used in verification decreases with the increasing lead time (Figs. 4, 7). TS, which measures the degree of coincidence between the forecast and observed events for a given precipitation threshold, is used here to indicate the relative forecast accuracy. For heavy precipitation no less than 25 mm day−1 during the 16 HPEs in June–July 2020, the TSs achieved by KISAM, KISAM_ms, and ECMWF ensemble mean generally decline with the extension of lead time (Figs. 4a, b). This point is also evidenced by the more occurrences of no skill (TS amounts to zero) at longer lead times (Figs. 4c, d). Noteworthily, the skill of ECMWF ensemble mean in forecasting heavy precipitation decreases more sharply than KISAM and KISAM_ms with the increase of lead time, especially between the lead time of 5 and 6 days, which is also demonstrated in Fig. 5. Exactly at the lead time of 5–6 days, advantages of KISAM and KISAM_ms over ECMWF ensemble mean in forecasting heavy precipitation emerge and get more highlighted at longer lead times (Figs. 4a, b). Through the comparison of performance in the forecast of each HPE, ECMWF ensemble mean beats KISAM and KISAM_ms for more times at 1–4 lead days. However, at lead times longer than 4 days, KISAM and KISAM_ms exhibit higher skill than ECMWF ensemble mean in more HPEs (Figs. 4c, d). All the above facts indicate that forecast products from KISAM perform better than ECMWF in capturing the precipitation at the threshold of 25 mm day−1 during the 16 HPEs at the medium-to-long range. Since TS does not distinguish the source of the forecast error, SR, POD, and BIAS are further examined.

    Figure 4.  Average TSs of (a) KISAM and ECMWF ensemble mean, and (c) frequencies of when KISAM and ECMWF ensemble mean score higher than each other or both score zero for the forecasting of daily precipitation ≥ 25 mm day−1 during the 16 HPEs at 1–15 lead days. (b) and (d) are the same as (a) and (c), but for KISAM_ms. The numbers of HPEs used in the verification on each lead day are indicated on the top axis(请作者再次确认大图题及各个小图的纵坐标名称及是否有单位).

    Figure 5.  Categorical performance diagram for KISAM and ECMWF ensemble mean in forecasting precipitation no less than 25 mm day−1 at the lead times of 1–15 days during the 16 HPEs. The horizontal and vertical axes indicate SR and POD, respectively. Blue dashed lines represent BIAS with labels on the outward extension of the line, while labeled yellow solid contours are TSs(请作者确认图中重叠数字,是否妥当).

    The four metrics including POD, FAR, BIAS, and TS are geometrically related and consequently can be plotted in a single diagram. In this visual representation, high and low TS could be easily traced back to the number of misses and false alarms (Roebber, 2009). Like that shown in TS(请问是宾语从句么?如果是,不能用that,改成As shown in TS?), POD and SR also exhibit a general decrease with the increasing lead time for both KISAM and ECMWF ensemble mean (Fig. 5). For the forecasting of precipitation at the 25 mm day−1 threshold during the 16 HPEs, the dichotomous forecast scores achieved by KISAM appear to be more concentrated than ECMWF ensemble mean. This demonstrates the above-mentioned point that the skill of KISAM decreases more slowly than ECMWF ensemble mean (Fig. 5). ECMWF ensemble mean tends to underforecast the frequency of heavy precipitation (no less than 25 mm day−1) and the degree of underforecasting gets higher with the increase of lead time. By contrast, KISAM yields a more accurate ratio of the heavy precipitation frequency to that in observation, with a slight underestimation at lead times of 6–12 days. Underforecasting means issuing fewer events at the threshold of 25 mm day−1 compared to observation, hence the false alarms rate (FAR, equivalent to one minus SR) could be reduced. However, like a coin has two sides, a high degree of the underforecasting results in a large fraction of misses and downgrades the POD. At lead times shorter than 4 days, a good balance between high POD and low FAR is achieved by the forecasts from ECMWF ensemble mean, leading to its better performance than KISAM. At 4–5 lead days, the advantage of ECMWF ensemble mean over KISAM is purely resulted from the low FAR, since its POD has declined to be matchable to that of KISAM. For the forecasts of ECMWF ensemble mean at the medium-to-long forecast range, a large fraction of misses (low POD) overwhelm the FAR as low as that in KISAM, which partly explains its deficiency in terms of TS.

  • The accurate depiction of the rainband for an HPE at the medium-to-long forecast range is critical for effective deployment of flood prevention strategies. Verification of the forecasts for the meridional location of the rainband could provide forecasters with useful information to most effectively employ the forecasts and make a subjective adjustment. Taking two HPEs as an example, the forecasted latitudinal location of the precipitation accumulation is firstly examined against observation in a visual way (Fig. 6). One took place during 21–25 June (first event) and the other occurred over 4–9 July (second event). As shown in Fig. 6, the meridional distribution of the accumulated precipitation in observation and forecasts at increasing forecast lead days are placed successively along the horizontal axis. For the zonal-averaged precipitation accumulation above 50 mm during the two HPEs, observations are both located between 27° and 32°N. Forecasts from ECMWF, KISAM, and KISAM_ms overestimate the precipitation intensity of the first event while underestimating that of the second one.

    Figure 6.  Latitudinal location of the observed zonal-mean accumulated precipitation (mm) averaged over 112°–121°E and the corresponding forecasts by (a, b) ECMWF, (c, d) KISAM and (e, f) KISAM_ms for the two HPEs during (left column) 21–25 June at 1–11 lead days and (right column) 4–9 July at 1–10 lead days.

    Regarding the first event, forecasts from ECMWF ensemble mean basically well depict the latitudinal location of heavy rainfall at the first six lead days but show notable southward deviation afterward (Fig. 6a). Comparatively, forecasts from KSIAM, whether the ensemble mean or the most similar one, show better performance in capturing the meridional location of heavy precipitation at 7–10 lead days (Figs. 6c, e). For the second event, northward displacement is found in the forecasts from ECMWF ensemble mean at almost all the lead times, which is noticeable at longer lead times (Fig. 6b). Forecasts from KISAM and KISAM_ms also place the heavy rainfall center northward than the observation. However, the displacement error in the forecasts from KISAM and KISAM_ms is overall smaller than that in the forecasts from ECMWF ensemble mean, especially for KISAM_ms (Figs. 6d, f). It could be therefore concluded that forecast products from KISAM possess higher skill than ECMWF ensemble mean in forecasting the meridional location of heavy rainfall for the two selected HPEs.

    The performance of KISAM and ECMWF ensemble mean in locating heavy precipitation center is further quantitively verified for all the 16 HPEs. The mean absolute error in forecasting the latitudinal location of the heaviest center for zonal-mean precipitation accumulation is illustrated in Fig. 7. The displacement errors increase remarkably with the extension of lead time for both ECMWF ensemble mean and KISAM. This indicates that forecasting of the latitudinal location of the heavy precipitation center at the medium-to-long forecast range is the most challenging. In comparison, the performance of KISAM and ECMWF ensemble mean in locating heavy precipitation center is matchable at shorter lead times. At 8–12 lead days, smaller displacement errors in KISAM show that KISAM better captures the latitudinal location of HPEs than ECMWF ensemble mean. This also indicates that the advantage of KISAM over ECMWF ensemble mean in TS at 8–12 lead days partly originates from its better performance in capturing the location of the rainband. At lead times longer than 12 days, considering the comparable displacement error in locating heavy precipitation, higher TS achieved by KISAM than ECMWF ensemble mean is inferred to be mainly caused by the significant underestimation of the heavy precipitation in ECMWF ensemble mean.

    Figure 7.  Mean absolute error in the forecasts from ECMWF ensemble mean and KISAM at 1–15 lead days for the latitudinal location of the heaviest center of the zonal-mean precipitation accumulation along 112°–121°E during the 16 HPEs. The numbers of HPEs used in the verification on each lead day are indicated on the top axis.

5.   Conclusions and outlook
  • In this study, the capability of a newly refined analog forecast model (KISAM) and an NWP model from ECMWF in forecasting the attention-attracting frequent occurrence of heavy precipitation events (HPEs) over YHRV in June–July 2020 is assessed. A total of 16 HPEs over YHRV are identified via an object-oriented method and are further used for verification. The verification is conducted for the heavy precipitation at the threshold of 25 mm day−1 based on several dichotomous metrics such as TS, POD, SR, and BIAS. In addition, the forecast skill in the latitudinal location of the zonal-mean precipitation accumulation is also evaluated. The main conclusions from this study can be summarized as follows:

    KISAM is re-optimized by using the meteorological archive containing 730 total days of HPEs during 1981–2015 to extend its applicability to both the long-persistent HPEs (no less than 3 days) and those with short duration (1–2 days). For the 44 HPEs in 2016–2018 remained for independent validation, the added value of KISAM compared to ECMWF ensemble mean for forecasting precipitation no less than 25 mm day−1 is confirmed at lead times longer than 4 days.

    In the operational forecasting context, the forecast bias in KISAM for heavy precipitation during the 16 HPEs increases with the forecast lead time, similar to that is seen in ECMWF ensemble mean(请问是想用宾语从句么,若是,不能用that引导,建议改成what). Notably, the skill of KISAM decreases more slowly than ECMWF ensemble mean in terms of the dichotomous scores. The advantage of KISAM over ECMWF for correctly forecasting heavy precipitation at the threshold of 25 mm day−1 emerges at the lead times longer than 5 days. Additionally, KISAM better captures the frequency of the observed heavy precipitation at the entire forecast range. In contrast, forecasts from ECMWF ensemble mean are prone to underestimate the frequency of heavy precipitation. The low FARs benefited from issuing fewer “yes” events contribute to a large part of the advantage of ECMWF ensemble mean over KISAM at the first five lead days. However, at the medium-to-long forecast range, the high degree of underforecasting from ECMWF leads to a large fraction of misses and low PODs, which accounts for a portion of its poorer performance compared to KISAM.

    For the latitudinal location of zonal-mean precipitation accumulation during the 16 HPEs, KISAM provides more accurate forecasts than ECMWF ensemble mean at 8–12 lead days. This also contributes a part to the higher TSs in the forecasts from KISAM at this forecast range. At lead times longer than 12 days, the higher TS achieved by KISAM is inferred to be mainly caused by the significant underestimation of heavy precipitation in ECMWF ensemble mean.

    In summary, the most valuable part of the forecast products from KISAM for HPEs over YHRV lies in the medium-to-long forecast range. At this lead time range, the advantage of KISAM is mainly due to its better depiction of the observed frequency of heavy precipitation at all lead times and the better capture of the latitudinal location of the zonal-mean precipitation accumulation at 8–12 lead days. From a future perspective, KISAM could still be probably improved, especially at shorter lead times. Introducing other new predictors closely linked to local physical processes governing the HPEs may be a way ahead. The candidates which have been used in several studies to improve the performance of analog models include the vertical velocity and thermodynamic variables such as instability indices and Convective Available Potential Energy (Gibergans-Báguena and Llasat, 2007; Mehrotra et al., 2014; Ben Daoud et al., 2016). Vertical motion is considered as a good predictor for small-scale convective systems and its meridional location shows good correspondence to that of heavy precipitation (Zhou et al., 2018). Testing of incorporating vertical velocity in the predictors of an analog model developed for the precipitation forecast over large river basins in France has been performed. However, the added value for the precipitation forecast could only preserve until the second lead day. The increase in skill is neutralized or even reversed by the inability of the ECMWF model to correctly forecast the vertical velocity at longer lead times (Ben Daoud et al., 2016). Therefore, how to choose the most pertinent new predictors depicting local-scale physical processes and strike a balance between the benefit and hindrance merit further study.

Reference (35)

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return