FEBRUARY 2018 Z H O U E T A L . 221

Comparative Assessment of Two Objective Forecast Models for Cases of Persistent Extreme Precipitation Events in the Yangtze–Huai River Valley in Summer 2016

BAIQUAN ZHOU State Key Laboratory of Severe Weather, Chinese Academy of Meteorological Sciences, China Meteorological Administration, and University of Chinese Academy of Sciences, Beijing, China

PANMAO ZHAI State Key Laboratory of Severe Weather, Chinese Academy of Meteorological Sciences, China Meteorological Administration, Beijing, China

RUOYUN NIU National Meteorological Center, China Meteorological Administration, Beijing, China

(Manuscript received 21 March 2017, in final form 16 December 2017)

ABSTRACT

Two persistent extreme precipitation events (PEPEs) that caused severe flooding in the Yangtze–Huai River valley in summer 2016 presented a significant challenge to operational forecasters. To provide fore- casters with useful references, the capacity of two objective forecast models in predicting these two PEPEs is investigated. The objective models include a numerical weather prediction (NWP) model from the European Centre for Medium-Range Weather Forecasts (ECMWF), and a statistical downscaling model, the Key In- fluential Systems Based Analog Model (KISAM). Results show that the ECMWF ensemble provides a skillful 2 spectrum of solutions for determining the location of the daily heavy precipitation ($25 mm day 1) during the PEPEs, despite its general underestimation of heavy precipitation. For lead times longer than 3 days, KISAM outperforms the ensemble mean and nearly one-half or more of all the ensemble members of ECMWF. Moreover, at longer lead times, KISAM generally performs better in reproducing the meridional location of accumulated rainfall over the two PEPEs compared to the ECMWF ensemble mean and the control run. Further verification of the vertical velocity that affects the production of heavy rainfall in ECMWF and KISAM implies the quality of the depiction of ascending motion during the PEPEs has a dominating influence on the models’ performance in predicting the meridional location of the PEPEs at all lead times. The su- periority of KISAM indicates that statistical downscaling techniques are effective in alleviating the deficiency of global NWP models for PEPE forecasts in the medium range of 4–10 days.

1. Introduction Lu 2000). However, accurately predicting PEPEs remains a challenge for forecasters in operational meteorology (Ebert Persistent extreme precipitation (PEP) has attracted a et al. 2003). The problem they face is primarily the in- great deal of attention because of its long duration and sufficient skill of operational global numerical weather association with natural disasters. A regional PEP event prediction (NWP) models to correctly predict the location, (PEPE) is capable of inducing severe flooding of rivers and intensity, and duration of PEPEs (Ralph et al. 2010; streams, which may threaten people’s lives and property Sukovich et al. 2014). Numerous evaluation studies have and cause significant disruption to regional economies been carried out on forecasts of precipitation and the re- (Galarneau et al. 2012). The two historical flood events that lated large-scale atmospheric circulation produced by the struck central eastern China in 1998 and Pakistan in 2010 NWP models from The Observing System Research and are typical examples that were triggered by clusters Predictability Experiment Interactive Grand Global of PEPEs (Galarneau et al. 2012; Lau and Kim 2012; Ensemble (TIGGE). With regard to heavy precipita- tion, poor performance of low-resolution global NWP Corresponding author: Panmao Zhai, [email protected] models with general underestimation has been reported in

DOI: 10.1175/WAF-D-17-0039.1 Ó 2018 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses). Unauthenticated | Downloaded 10/04/21 05:24 PM UTC 222 WEATHER AND FORECASTING VOLUME 33 many studies (Krishnamurti et al. 2009; Ralph et al. 2010; 2016; Ben Daoud et al. 2016). Considering the performance Willems et al. 2012; Zhi et al. 2013; Su et al. 2014; Sharma of ECMWF is often better than other NWP models (Park et al. 2017). For an individual extreme precipitation et al. 2008; Niu and Zhai 2013; Zhou et al. 2015; Swinbank event, Kobold and Suselj (2005) found that the European et al. 2016), Zhou and Zhai (2016) used large-scale circu- Centre for Medium-Range Weather Forecasts (ECMWF) lation variables derived from the ECMWF ensemble mean underestimated the 27–28 June 1997 precipitation in a as predictors to establish a new statistical downscaling Slovenian catchment by 60%. Research conducted by model called the Key Influential Systems Based Analog Lavers and Villarini (2013) indicated that ensemble mean Model (KISAM) for PEPE forecasts in the YHRV. Be- rainfall patterns from NWP models in TIGGE initialized cause of the superior skill of KISAM in reforecasting, we at a roughly 4-day lead time failed to forecast the persistent are motivated to investigate the performance of KISAM nature of a PEPE. Characterized by an extreme and per- and ECMWF in real-time operational forecasts of PEPEs. sistent nature, PEPEs pose enormous difficulties to oper- Two PEPEs that occurred in the YHRV during ational forecasters and NWP models. There is thus an urgent summer 2016 and met the criteria specified by Chen and need for effective forecast methods for PEPEs, for disaster Zhai (2013) were selected for our performance assess- prevention and mitigation, particularly across regions like ment. Operational forecast products based on NWP the Yangtze–Huai River valley (YHRV), which is one of the models failed to correctly capture the meridional loca- most important agricultural zones and densely populated tions of the rainbands of these two PEPEs, which re- regions in China (Ralph et al. 2005; Lu 2000). Within the sulted in an ineffective deployment of flood prevention context of global warming, with more intense and frequent strategies. To improve the performance of the NWP extreme precipitation events projected (Hartmann et al. models for PEPEs, forecast verification is crucial 2013; Hirsch and Archfield 2015), the YHRV will be more (Sukovich et al. 2014). Therefore, this study aims to likely to suffer severe socioeconomic damage due to the verify the performance of ECMWF, including all flooding and landslides caused by PEPEs. ensemble members and its ensemble mean, as well as Helpfully, studies on PEPEs in China are now KISAM, for PEPE forecasts. Their strengths and emerging. Instead of investigating extreme precipitation weaknesses will be analyzed to provide forecasters with primarily on a daily basis (e.g., Karl and Knight 1998; useful guidelines to employ the NWP models and sta- Zhai et al. 2005), Chen and Zhai (2013) introduced a tistical downscaling models effectively. Furthermore, new definition for regional PEPEs based on the ex- through comparison of the atmospheric circulation and tremity, temporal persistence, and spatial contiguity of physical conditions affecting the process of precipitation the precipitation. The criteria for this PEPE definition production in ECMWF and KISAM, we will further include 1) that daily precipitation amounts of not less address the leading factors behind their advantages and than 50 mm should be observed by at least three disadvantages in PEPE forecasts. neighboring stations and 2) that the condition in the first The data employed and the methodology are de- criterion should persist for at least 3 days at every sta- scribed in section 2, followed by an introduction of the tion. The PEPE comes to an end when the first criterion characteristics of the two PEPEs and corresponding is not satisfied for the following two consecutive days. In circulations in section 3. A comparative assessment of their subsequent research, they reported that PEPEs the two objective forecasts for the two PEPEs is given in typically result from concurrent combinations of per- section 4. Section 5 presents an analysis of the superi- sistent anomalies from the displacement of the South ority of KISAM for longer lead times, with conclusions Asian high and jets, the development of blocking pat- and further discussion provided in section 6. terns, and a westward shift of the western Pacific sub- tropical high (WPSH) (Chen and Zhai 2014). Therefore, utilizing the revealed circulation patterns to downscale 2. Data and methodology PEPEs is a sensible strategy to alleviate the deficiency of a. Data global NWP models for extreme precipitation forecasts, since they exhibit better skill in forecasting large-scale The observed daily precipitation dataset from about circulation variables and key synoptic systems than ex- 2200 stations covering China in summer 2016 is provided treme precipitation (Park et al. 2008; Pelly and Hoskins by the National Meteorological Center of the China 2003; Niu and Zhai 2013; Niu et al. 2015; Zhou et al. Meteorological Administration (CMA). Daily atmo- 2015). Statistical downscaling models are usually de- spheric reanalysis data from the ERA-Interim dataset, veloped and employed for weather and climate pre- including wind vectors, geopotential height, and relative diction (Wilby and Wigley 1997; Zorita and Von Storch humidity in the summers of 1981–2010 and 2016, with a 1999; Fan et al. 2008; Wang and Fan 2009; Caillouet et al. horizontal resolution of 18318, are used to depict the

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC FEBRUARY 2018 Z H O U E T A L . 223

FIG. 1. Schematic diagram for the KISAM downscaling process. climatology and circulation features during summer 2016, also comprise the same variables that are obtained from respectively. ERA-Interim is the latest global atmospheric the NCEP–NCAR reanalysis. The forecasts of vertical analysis produced by the ECMWF and is based on the 2006 velocity at 500 hPa from the ECMWF ensemble mean in version of the integrated forecasting system (Cy31r2; Dee the medium range of up to 11 days, with a forecast step et al. 2011). Considering KISAM is based on an analog of 6 h, initialized on each day from 1 June to 6 July at a method, daily precipitation data from 756 weather stations 28328 horizontal resolution, are also used. over China during 1951–2010 from the National Meteoro- For ease of comparison, the station precipitation data logical Information Center of the CMA are used to obtain are interpolated onto an identical grid of 18318 through the precipitation for the historical PEPE records. Re- the Cressman interpolation method, and circulation analysis data from the National Centers for Environmental variables are interpolated onto an identical grid of 2.583 Prediction–National Center for Atmospheric Research 2.58 through bilinear interpolation. (NCEP–NCAR; Kalnay et al. 1996) serve as the historical b. Brief introduction to KISAM pool of large-scale circulation for matching analogs. The daily large-scale variables in the summers of 1951–2010, KISAM is trained and verified using the 25 PEPEs with a horizontal resolution of 2.5832.58,fromtheNCEP– that Chen and Zhai (2013) identified in the YHRV NCAR reanalysis, comprise the zonal wind at 200 hPa, (268–348N, 1128–1218E) over 1951–2010. The schematic geopotential height and vertical velocity at 500 hPa, and process of KISAM is displayed in Fig. 1. Based on the zonal and meridional water vapor transport at 700 hPa. circulation patterns from upper to lower levels condu- The predicted large-scale circulation variables and cive to the PEPEs identified by Chen and Zhai (2014), total precipitation from the ECMWF ensemble, KISAM selects four large-scale atmospheric variables alongside a control run, are retrieved from initializations from the circulation patterns as predictors, including at 1200 UTC on each day from 1 June to 6 July in zonal wind at 200 hPa, geopotential height at 500 hPa, summer 2016, for lead times of up to 15 days with a and zonal and meridional water vapor transport at forecast step of 1 day. The control forecast, a single 700 hPa. The forecast data for the predictors used by member of the ECMWF ensemble, is usually an un- KISAM are the large-scale circulation variables ob- perturbed analysis generated by a data assimilation tained directly from ECMWF at lead times of 1–15 days. procedure. The ensemble-mean forecasts are derived by The similarity metric, with the calculation shown in the equally weighting all ensemble members. The horizon- orange box in Fig. 1, is the weight-assigned cosine of the tal resolution of the total precipitation is 18318. The angle between two patterns. The variables A and B large-scale circulation variables obtained from the represent two atmospheric fields, and n is the total ECMWF ensemble at a 28328 horizontal resolution number of grid points. The weight function G(i) is the

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC 224 WEATHER AND FORECASTING VOLUME 33 normalized absolute correlation coefficient over the KISAM are performed against the NCEP–NCAR re- study region between each predictor at each grid point analysis in summer 2016. Vertical velocity is not used by and the area-averaged precipitation over the training KISAM as a predictor and is not involved in the period. The weights assigned can help emphasize the downscaling process. Thus, the vertical velocity exam- role in matching analogs of those key large-scale circu- ined in KISAM is the average vertical velocity obtained lation systems that have larger correlations with the from the NCEP–NCAR reanalysis of the three most PEPEs. The similarity is measured between forecasts analogous historical records. We use the equitable and the records in the historical pool for the selected threat score (ETS) to measure the fraction of observed four predictors individually, and the resulting four sim- events that are correctly predicted, and it accounts for ilarities S200, S500, S700qu, and S700qv are then combined to hits associated with random chance (Wang 2014). The obtain one integrated similarity score. As the equation ETS ranges from 21/3 to 1, with 1 indicating a perfect in the right-hand box in the third row of Fig. 1 illustrates, score. It is defined as the normalized weights P200, P500, P700qu, and P700qv are hits 2 hits assigned to the four similarities, respectively, to get the 5 random ETS 1 1 2 and combined similarity. Another parameter, Pcv,in- hits misses false alarms hitsrandom troduced in KISAM, is a critical value for judging (hits 1 misses)(hits 1 false alarms) hits 5 , whether a historical record is sufficiently analogous to random hits1misses1false alarms1correct negtives the target day. These parameters are determined by an optimization algorithm known as a ‘‘cuckoo where hits, misses, false alarms, and correct negatives search’’ in the training process to optimize KISAM. are the four combinations of forecasts (yes or no) and The optimization is designed to seek one set of pa- observations (yes or no), called the joint distribution. rameters (P , P , P , P ,andP )tomax- 200 500 700qu 700qv cv The categorical bias score (BS) is used to measure the imize the threat score of the heavy precipitation forecast performance for the frequency of precipitation forecast for the historical PEPE days and give KI- exceeding different thresholds. It is calculated as SAM the best capability to discriminate PEPE days from non-PEPE days. The parameters obtained after hits 1 false alarms BS 5 . training show that the weights assigned generally hits 1 misses show a decrease from lower to upper levels. This suggests that the lower layer has a more crucial in- The perfect score of BS is 1. Rainfall in eastern China is fluence on the analog scheme. When the obtained usually concentrated in a zonal belt and moves pro- integrated similarity between a historical record and gressively toward the north along with the summer the target day is larger than Pcv, KISAM identifies the monsoon advancing northward (Tao 1980; Ding 1992; historical record as an analog. Circulation and phys- Samel et al. 1999). Therefore, the meridional location of ical conditions such as vertical motion on these the rainband is an important measure when evaluating a analogous days actually control the precipitation that model’s PEPE forecast performance. To indicate the KISAM produces. In the end, if KISAM successfully direction of deviations of the meridional location, the finds three or more analogs, it will issue a pre- zonal-mean accumulated precipitation over 1128–1218E cipitation forecast using the weighted average of the during PEPEs is investigated. Moreover, the pre- three most similar analogs. The weights are the sim- cipitation root-mean-square error (RMSE) is calculated ilarity scores of the analogs. Otherwise, KISAM will to quantitatively examine the average magnitude of the not produce a precipitation forecast and assigns forecast errors for the PEPE forecast. It is noted that missing values for the precipitation field. More spe- when KISAM fails to issue a precipitation forecast, the cific details of KISAM are described by Zhou and constant precipitation field of missing values will be Zhai (2016). taken as zero when calculating the RMSE. Besides ver- ification of PEPE forecasts from ECMWF and KISAM, the c. Analysis method investigation will focus on the key large-scale circulation and Verification of the precipitation forecasts from the physical processes that are critical for the production of ECMWF ensemble members and ensemble mean, as precipitation and subsequently affect PEPE fore- well as from KISAM, is conducted with respect to the cast accuracy in these two models. When verifying observed precipitation collected from about 2200 sta- daily precipitation during a PEPE, we aim to investi- tions geographically covering China during summer gate the average performance of a single 15-day forecast 2016. Verifications of vertical velocity at 500 hPa leading initialized on one selected lead day for the whole to the production of heavy precipitation in ECMWF and PEPE process, rather than examine the mean performance

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC FEBRUARY 2018 Z H O U E T A L . 225

FIG. 2. Accumulated precipitation (mm) of all the stations in the YHRV (1128–1218E, 268–348N) during the two PEPEs in summer 2016: (a) event 1 (18–22 Jun), (b) event 2 (1–6 Jul), and (c) time–latitude cross section of daily precipitation 2 (mm day 1) averaged along 1128–1218E from 1 Jun to 15 Jul. In (a) and (b), gray shading indicates the terrain of the region (m), colored dots represent the amount of precipitation at each station, and the orange line denotes the Yangtze River. of multiple 15-day forecasts with the same forecast length (Tao 1980). During this period, PEPEs are prone to corresponding to each day of a PEPE. Therefore, forecast occur, with cold and dry air persistently encountering lead time for a PEPE in this paper refers to the number of anomalous warm and moist air in the same region (Ding days in advance of the first day of the PEPE. Taking the first and Reiter 1982; Chen and Zhai 2014). Two PEPEs in PEPE (18–22 June 2016) as an example, a 3-day lead time summer 2016 that induced severe flooding in the YHRV means the forecast is initialized on 15 June. In this case, the are chosen as our verification cases; one occurred during forecasts for 19–22 June are in fact at lead times of 4–7 days. 18–22 June (hereafter event 1), and the other took place Thus, for a 15-day forecast, a 5-day PEPE can be forecasted over 1–6 July (hereafter event 2). From the distribution as early as 11 days in advance. Following Zhou et al. (2015) of the observed accumulated precipitation during these and Niu and Zhai (2013), indices for the meridional location two PEPEs in Figs. 2a and 2b, it can be seen that, for of the WPSH and the East Asian subtropical westerly jet both events, the rainbands are generally located along (EASWJ) are calculated. The average axis of the EASWJ is the Yangtze River. The stations with accumulated defined as the average of the latitude with maximum zonal rainfall greater than 100 mm are mainly concentrated 2 wind above 30 m s 1 at 200 hPa at each longitude over 1108– between 298 and 338N for event 1, as shown in Fig. 2a. 1308E . The mean position of the WPSH ridge is de- Compared to event 1, the stations with heavy rainfall termined as the mean latitudinal position of the high pres- accumulation in event 2 appear to be more concentrated sure ridge (u 5 0and›u/›y , 0, where u represents zonal along the Yangtze River. The two PEPEs are both very wind) surrounded by the 588-dagpm contour at 500 hPa to strong, with the maximum accumulated precipitation the north of 108Nover1108–1308E. reaching 200–300 mm, and above 500 mm, respectively. According to official statistics, event 2 was the most in- tense precipitation event in 2016 in China. Severe 3. Characteristics of the two PEPEs and flooding related to these two PEPEs struck the lower corresponding circulations reaches of the Yangtze River, resulting in substantial With the summer monsoon shifting northward, the damage to local agriculture. From the time–latitude rainband in eastern China shifts to the YHRV in June cross section of daily precipitation along 1128–1218E

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC 226 WEATHER AND FORECASTING VOLUME 33

FIG. 3. Averaged atmospheric circulation patterns at different levels during the two PEPEs showing (left) event 1 and (right) event 2. Among them, results are shown for (a1),(a2) 200, (b1),(b2) 500, and (c1),(c2) 700 hPa. The blue lines in all panels represent the Yangtze River and the Yellow River. At 200 and 500 hPa, black lines represent geopotential height contours, and red lines indicate climatological geopotential height $ 1252 dagpm west of 1508E at 200 hPa and .588 dagpm east of 908E at 500 hPa, over 1981–2010. Shading at 200 hPa indicates the magnitude of 2 the westerly wind $ 30 m s 1, whereas at 500 hPa it represents the anomaly between geopotential height during the 2 PEPEs and the climatology over 1981–2010. At 700 hPa, vectors indicate wind with magnitude $ 6ms 1, and blue 2 shading represents anomalies between specific humidity during the PEPEs and the climatology over 1981–2010 (g g 1).

(Fig. 2c), it is noted that the rainband shifts from south to YHRV is located at the right-hand entrance of the jet north during event 1, while it remains relatively steady stream above the Sea of Japan. Compared to the cli- during event 2. It can also be clearly seen that the rainfall matology over 1981–2010, the South Asia high is more amount in event 1 is relatively low compared to event 2 intense and shifts eastward. These conditions both pro- (Fig. 2c). vide strong upper-level divergence over the YHRV, As indicated in Fig. 3, the large-scale atmospheric which is favorable to the occurrence of a PEPE (Samel circulations of the two PEPEs are both characterized by et al. 1999; Zhu et al. 2010). At 500 hPa, a pair of the typical double-blocking high circulation patterns blocking highs is located over the Ural Mountains and described by Chen and Zhai (2014). At 200 hPa, in the Sea of Okhotsk, respectively (Figs. 3b1 and 3b2). Figs. 3a1 and 3a2, a subtropical jet stream is located to The WPSH shifts westward and produces an anoma- the north of the YHRV and near the Sea of Japan. The lous anticyclone with southwesterly flow transporting

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC FEBRUARY 2018 Z H O U E T A L . 227 excessive moisture into the YHRV during the event one-half or more of the ECMWF ensemble members compared to the climatology (Figs. 3c1 and 3c2). These have skill in locating heavy rainfall during the process of atmospheric conditions all match the double-blocking the two events. It is noted that the ECMWF and KISAM high circulation pattern and the associated synoptic ETSs are relatively low at longer lead times, which in- features reported by Chen and Zhai (2014), which play a dicates that PEPE forecasting is also a challenge for the significant role in the development and maintenance of ECMWF ensemble and KISAM. Thus, there is still a the two PEPEs. major need to improve the prediction of PEPE in con- sideration of the related catastrophic consequences. A comparison of KISAM and the ECMWF ensemble 4. Comparative assessment of the two objective results reveals that the skill of KISAM is lower at lead forecasts times of 1–3 days. However, for event 1 (Fig. 4a), the As a result of the northward progression of the rainfall ETSs of KISAM surpass the median of the ECMWF band in event 1 and the large amount of rainfall in event ensemble at lead times of 4–10 days. KISAM is superior 2, it is very difficult for forecasters to precisely predict to 75% or more of the members of ECMWF in scoring these extreme events. Objective forecasts from NWP or better ETSs at lead times of 5–9 days. It is noteworthy downscaling models have potential use for forecasters. that the KISAM forecast does a better job at capturing 2 Verification and comparison of forecasts from the the location of rainfall at the 25 mm day 1 threshold for ECMWF ensemble members and ensemble mean, as event 1 when the lead time increases from 1 to 8 days. well as KISAM, are conducted for the two PEPEs in a Similarly, the ETSs of the control forecast also show an medium forecast range of up to 10 days, to provide increase at lead times of 8 or more days compared to the useful information for the improvement of the PEPE middle-range forecast. This can be explained by the fact forecast. that verification is conducted against these two individ- Considering the general underestimation of global ual cases in this paper. However, the generally de- 2 NWP models for precipitation in excess of 25 mm day 1 creasing skill of the NWP models with increasing lead (Ralph et al. 2010; Su et al. 2014; Sharma et al. 2017), the time reported in the literature is achieved by verification performance of ECMWF and KISAM is examined using scores averaged over many days or seasons (Park et al. the ETS for daily precipitation at a threshold intensity 2008; Lavers and Villarini 2013). For these two cases, the 2 of 25 mm day 1 with respect to the observations. For location of the rainband is indeed more accurately pre- forecasts initialized on each day from 7 to 17 June and dicted at longer lead times. This point is evidenced by from 21 to 30 June for events 1 and 2 in the YHRV, the well-predicted meridional location of the rainband the ETS is first calculated for precipitation at the at longer lead times by KISAM (see Fig. 6c1, which will 2 25 mm day 1 threshold on all the days of each PEPE be discussed later). For event 2 (Fig. 4b), the skill of process, and then the ETSs are averaged over the whole KISAM is compatible with the median of the ECMWF PEPE process. Therefore, for a given lead day, multiday ensemble at 4–10 days, with exceptions at lead times of 5 forecasts corresponding to the whole PEPE process are and 7 days. In short, at lead times longer than 3 days, the used to achieve the average verification score. As in- accuracy of KISAM at least exceeds half of the ECMWF dicated by the average ETSs in Fig. 4, with the exception ensemble members in locating heavy precipitation dur- of KISAM at several lead times, KISAM and most ing the two PEPEs. ECMWF ensemble members show skill in detecting the To assess the quantitative accuracy of these forecasts location of precipitation at a threshold intensity of in predicting heavy daily rainfall, the RMSE for the 2 2 25 mm day 1 during these two PEPEs with a lead time of forecasting of precipitation at the 25 mm day 1 thresh- up to around 9 days. The ECMWF ensemble mean old is investigated (figure not shown). RMSE for pre- 2 shows an obvious advantage over KISAM at lead times cipitation above 25 mm day 1 is calculated between the 2 less than 4 days but performs poorly afterward, as in- precipitation observations that exceed 25 mm day 1 and dicated by ETS values nearly approaching 0. This is the corresponding forecasts. The ECMWF ensemble partly due to the smoothing effect by the ensemble av- mean has the advantage over most ensemble members, erage. In contrast, KISAM generally exhibits better skill including the control run. For event 1, KISAM shows than the ECMWF ensemble mean in the forecast range the best performance at lead times of 2–9 days. The 2 of 5–10 lead days, especially for event 1. Moreover, the performance of KISAM for event 2 at the 25 mm day 1 ETSs of all ECMWF ensemble members appear to de- threshold shows dramatic variation with increasing lead crease with increasing lead time. The median ETSs of time. However, with the exception of lead days 5 and 7, the ECMWF ensemble are all equal to or greater than the performance of KISAM can still match that of the zero for all lead times, which indicates that nearly ECMWF ensemble mean forecast for lead times longer

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC 228 WEATHER AND FORECASTING VOLUME 33

FIG. 4. Average ETSs of all ECMWF ensemble members, the ensemble mean, and KISAM, 2 for the forecasting of daily precipitation of no less than 25 mm day 1 during (a) event 1 at lead times of 1–11 days and (b) event 2 at lead times of 1–10 days. For a given lead day, multiday forecasts corresponding to the whole PEPE process are used to achieve the average ETSs. Boxplots show the ranges in average ETS spanned by 51 ECMWF ensemble members, the red line shows the median of the ensemble, the red plus signs indicate the outliers of the boxplot for the ECMWF ensemble, the red stars denote the ETSs of the KISAM forecasts, and the blue asterisks and green diamonds denote the ETSs of the ECMWF ensemble mean and control forecasts, respectively. than 3 days. To discriminate the forecast error of the performs well for the frequency of light precipita- 2 rainfall amount from location error, categorical bias tion ($1mmday 1), but shows a general tendency scores are explored for the ECMWF ensemble and to underpredict (BS , 1) heavy precipitation 2 KISAM at different lead times and precipitation thresh- ($25 mm day 1). For KISAM, consistent over- olds (Fig. 5). For event 1, KISAM overpredicts light estimation or underestimation is found for light and 2 precipitation ($1mmday 1) and heavy precipitation heavy precipitation. 2 ($25 mm day 1; Figs. 5a1 and 5b1). In contrast, the To determine the direction of forecast deviations for ECMWF ensemble performs better, with most members the meridional location of the entire PEPE process, achieving perfect bias scores for light precipitation. which the ETS and BS do not indicate, the forecasts of 2 However, the 25 mm day 1 precipitation is under- the zonal-mean accumulated precipitation during each predicted by the ECMWF ensemble at lead times of PEPE process averaged over 1128–1218E are verified for 1–8 days; also, at lead times longer than 8 days, about half lead times of 1–9 days. For each forecast initialization of the members overpredict heavy precipitation while the time, a single forecast of zonal-mean multiday total other half show an underestimation. For event 2, un- precipitation throughout the PEPE is used to match with derestimation of light and heavy precipitation is seen for the observed precipitation accumulation. Figure 6 shows both the ECMWF ensemble and KISAM, with the larger the meridional location of the precipitation accumula- error for heavy precipitation. It is also noted that the tion at lead times ranging from 9 days to 1 day, placed ECMWF ensemble mean shows significant under- successively from far left to right. The observed me- 2 prediction of the 25 mm day 1 precipitation frequency ridional locations of precipitation accumulated during due to the smoothing effect. Summarizing the results 18–22 June and during 1–6 July for events 1 and 2, re- from the bias scores, it seems that the ECMWF ensemble spectively, are placed on the far right. The discrepancy

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC FEBRUARY 2018 Z H O U E T A L . 229

FIG. 5. Average bias scores of daily precipitation forecasts at different thresholds during the PEPEs for all ensemble members, including the control run of ECMWF, the ensemble mean of ECMWF, and KISAM: (a1) event 2 2 2 1 at the 1 mm day 1 threshold, (b1) event 1 at the 25 mm day 1 threshold, (a2) event 2 at the 1 mm day 1 threshold, 2 and (b2) event 2 at the 25 mm day 1 threshold. The legend in (b2) is applicable to all panels; the pink dashed lines denote the perfect bias score. between the forecasts at all lead times and the obser- underestimation (particularly at 3–6 days) and dis- vation clearly indicates a bias in predicting the latitude placement error of heavy rainfall for longer lead times of the zonal-mean heavy rainfall center during the (Fig. 6b1). The rainfall pattern is depicted well by KI- PEPE. The forecasts include the ECMWF ensemble SAM, albeit the amount is overestimated and the loca- mean, the ECMWF control forecast, and KISAM. Since tions of the maxima are shifted 28 southward, especially the ensemble mean represents the average locations of at lead times of 2–3 days (Fig. 6c1). For event 2, as shown precipitation centers depicted by all ensemble members, in Fig. 6a2, the precipitation maxima predicted by the verification is conducted for the ECMWF ensemble ECMWF ensemble mean gradually shifts to the north mean instead of all ECMWF ensemble members. The with increasing lead time, and reaches 348N compared to ECMWF control run is selected from the ECMWF en- the observed 318N at lead day 7. The same northward semble as an independent reference. For event 1, the deviation is found for the ECMWF control forecast observed zonal-mean precipitation accumulation above (Fig. 6b2). From Fig. 6c2, it is noted that the deviations 50 mm is located between 308 and 338N. Rainfall from of KISAM for forecasts of the meridional location of ECMWF ensemble mean was slightly underestimated event 2 at various lead times are not consistent. The during the entire forecast period and placed more forecast precipitation center is located about 28 north- northward than the observation, notably at lead times ward at lead times of 8–9 days, and slightly to the south greater than 4 days (Fig. 6a1). The ECMWF control run at lead times of 1–7 days (Zhou and Zhai 2016). A point better captures the precipitation location and amount to note is that all three forecasts predict weaker in- on the first two lead days, but shows considerable tensities of the accumulated precipitation for event 2.

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC 230 WEATHER AND FORECASTING VOLUME 33

FIG. 6. Meridional location of observed and forecasted zonal-mean accumulated precipitation (mm) averaged over 1128–1218E for the two PEPEs. Forecasts for lead times from 9 days to 1 day are placed successively from the far left to right, and the observation is shown on the far right. For each lead day, a single forecast of the multiday total precipitation throughout the PEPE is used. Shown are results for (left) event 1 and (right) event 2. (a1),(a2) Forecasts of the ECMWF ensemble mean, (b1),(b2) the ECMWF control forecasts, and (c1),(c2) KISAM forecasts.

Nonetheless, the ECMWF control forecast outperforms forecasting the frequency and amount of heavy pre- the other two forecasts. The weak precipitation accu- cipitation. With respect to the meridional location of the mulation of event 2 produced by KISAM is due to its two PEPE processes, KISAM forecasts are generally failure to produce forecasts for the first 2 days of event 2; closer to the observation at lead times longer than thus, the precipitation on these 2 days is set as missing 3 days, due to the significant northward divergence for values and not counted in the accumulation. forecasts of the ECMWF ensemble mean and the con- To sum up, KISAM is generally as good as the median trol forecast. KISAM uses the large-scale atmospheric of all the ensemble members of ECMWF at lead times variables from the ECMWF ensemble mean to down- longer than 3 days for detecting the location of heavy scale precipitation over the PEPE. Therefore, the fore- precipitation. In addition, the investigation based on the casts generated by KISAM and ECMWF are influenced RMSE and BS indicates that KISAM also performs as partly by the same large-scale atmospheric background. well as most ensemble members of the ECMWF in The underlying reason contributing to the different

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC FEBRUARY 2018 Z H O U E T A L . 231

TABLE 1. Dates of the three most similar analogs found by KISAM for event 1, at 9-day lead time. Boldface indicates that the day is among the cases under the double-blocking high pattern identified by Chen and Zhai (2014).

Objective day 18 Jun 19 Jun 20 Jun 21 Jun 22 Jun Analogous days 10 Jul 2003 18 Jun 1955 18 Jun 1955 1 Jul 1991 7 Jul 1954 matched 18 Jun 1955 10 Jul 2003 1 Jul 1991 15 Jun 1998 15 Jun 1998 12 Jun 1991 1 Jul 1991 16 Jun 1998 18 Jun 1955 1 Jul 1991 levels of forecast performance between KISAM and event 2 at 6 days is comparable to that at 3 days. With the ECMWF is expounded upon in the next section. lead time increasing to 9 days, the intensity and extent of the subtropical jet as predicted by ECMWF are both underestimated. For the WPSH, which is usually depicted 5. Analysis of the superiority of KISAM for longer by geopotential heights of 586 and 588 dagpm, the pre- lead times diction of event 1 shows good skill, with the extent and As Chen and Zhai (2014) reported, PEPEs typically intensity close to the observation at the lead time of result from concurrent anomalies of key large-scale 3days(Fig. 7b1). The predicted northern edge of the systems such as subtropical jets, blocking highs, and WPSH at 3 and 6 days is located farther north than ob- the WPSH. Moreover, the predictors of KISAM contain served. Interestingly, ECMWF describes the northern the large-scale variables obtained directly from edge of the WPSH more accurately at the lead time of ECMWF. For the spatial field of these predictors, the 9 days. For event 2, the extent of the WPSH is over- key large-scale systems are all assigned more weight in estimated slightly at the lead times of 3, 6, and 9 days the statistical downscaling scheme of KISAM, which (Fig. 7b2). The northwest quadrant of the WPSH was introduced in section 2b. Therefore, the skill of the deviates a little to the north at 3, 6, and 9 days for event 2. ECMWF ensemble mean in predicting these key large- However, the northward deviation of the northwest scale systems has significant influence on the perfor- quadrant of the WPSH at 9 days appears to be smaller mance of KISAM and the ECMWF ensemble for PEPE than that at 3 and 6 days. In addition, water vapor forecasts. The analysis is therefore focused on the skill of transport is predicted well for events 1 and 2, with the ECMWF ensemble mean in predicting these key abundant moisture advected to the YHRV (Figs. 7c1 large-scale systems during the two PEPEs. and 7c2). As mentioned above, KISAM is based on an analog The forecast errors of the ECMWF ensemble mean method and uses the large-scale circulation depicted in for the mean meridional position of the WPSH and the the ECMWF ensemble mean forecast to search for EASWJ over the region 1108–1308E on each day of the analogous circulations from the historical pool. Even at two PEPEs are quantified using indices of the average the lead time of 9 days for event 1, the three most similar WPSH ridge and EASWJ axis. Positive and negative analogs that KISAM finds from the historical pool (Table 1) errors indicate northward and southward deviations, are almost all from the PEPE cases under the double- respectively. From Fig. 8a1, it can be seen that the blocking high pattern identified by Chen and Zhai (2014). southward deviation of the forecast at a 6-day lead for Since the analogous circulations are characterized by the the average WPSH ridge increases with valid time dur- double-blocking high pattern, the synoptic patterns of ing event 1. The northward deviation of the average the predictors forecasted by the ECMWF ensemble WPSH ridge forecast at 9 days turns to a southward mean must also exhibit a double-blocking high pattern. deviation with time during event 2 (Fig. 8a2). With the This reflects indirectly that the large-scale circulation of exception of these two circumstances, the deviations of KISAM taken from the ECMWF ensemble mean well the WPSH ridge forecast at the lead times of 1, 3, 6, and reproduces the double-blocking high patterns during 9 days are relatively steady over the two PEPEs. For the these two events. In addition, the performance of the average EASWJ axis, the forecasts all locate this axis to ECMWF ensemble mean for predicting the subtropical the north of that observed at lead times of 1, 3, and jet and WPSH is examined. Figures 7a1 and 7a2 illustrate 6 days, with the deviation more evident at 6 days that, for both events, the subtropical jet is predicted well (Fig. 8b1). This behavior confirms the result shown in by ECMWF in terms of extent, intensity, and location at Fig. 7a1. For event 2, forecasts of the average EASWJ the lead time of 3 days. At 6 days, the jet axis deviates axis from the ECMWF ensemble mean show general significantly to the north at 908–1208E during event 1. The southward deviations at the lead time of 1 day. The di- performance of ECMWF for the subtropical jet during rection of the EASWJ axis forecast deviation shifts from

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC 232 WEATHER AND FORECASTING VOLUME 33

FIG. 7. Prediction of the ECMWF ensemble mean and observations for key large-scale systems during events 1 and 2. (a1),(a2) The EASWJ at 200 hPa, (b1),(b2) the WPSH at 500 hPa, and (c1),(c2) the water vapor transport at 700 hPa. The blue shading indicates the observed EASWJ and WPSH. Shown are results for (left) event 1 and (right) event 2. The predictions are all presented as contours at lead times of 3, 6, and 9 days. At 200 hPa, zonal 2 winds $ 30 m s 1 are displayed. At 500 hPa, heights of 586 and 588 dagpm are shown. Water vapor transport at 700 hPa is only shown for a lead time of 6 days; blue shading indicates specific humidity and vectors indicate wind 2 with magnitude $ 6ms 1. The orange lines denote the Yellow and Yangtze Rivers. northward to southward during event 2 at 3 and 6 days. (Zhou et al. 2015), the MAE of the EASWJ axis aver- Deviations of the EASWJ axis forecast at 9 days oscil- aged over 1008–1258E can reach up to around 48 at late around zero and are generally southward. The mean the lead time of 15 days. The MAEs of the two in- absolute error (MAE) of the WPSH ridge forecast av- dexes in these two studies are computed over many eraged daily during each PEPE is computed for all lead summer days. times. The maximum MAE among all lead times for the Thus, for the two PEPE cases, the bias of the ECMWF average WPSH ridge amounts to 1.18 and 1.38 for events ensemble mean in predicting the meridional location of 1 and 2, respectively. Similarly, the maximum MAEs for the WPSH and the EASWJ up to 9 days ahead of the the average EASWJ axis reach 48 and 2.78 for events 1 PEPE is smaller than the average error reported in the and 2, respectively. As Niu and Zhai (2013) reported, literature (Niu and Zhai 2013; Zhou et al. 2015). With the MAE of the mean position of the WPSH ridge over increasing lead time, the forecast deviations of the in- 1108–122.58E is as large as about 3.58 at lead times of up tensity and meridional position of the key large-scale to 15 days. Based on our previous work related to the systems like the WPSH and the EASWJ show poor as- evaluation of the EASWJ prediction in the NWP models sociation with that of the PEPEs for ECMWF. For

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC FEBRUARY 2018 Z H O U E T A L . 233

FIG. 8. Forecast error (8) of the ECMWF ensemble mean forecast for the mean position of the WPSH ridge and EASWJ axis over 1108–1308E at lead times of 1, 3, 6, and 9 days: (a1) event 1 for the average WPSH ridge, (b1) event 1 for the average EASWJ axis, (a2) event 2 for the average WPSH ridge, and (b2) event 2 for the average EASWJ axis. The legend in (a1) is applicable to all panels. example, the performance of the ECMWF ensemble precipitation. The effect of long duration mainly mani- mean in depicting the WPSH and the EASWJ is com- fests in the intensity of the precipitation accumulation, parable at lead times of 1, 3, and 6 days for event 2. as verified and discussed above. In addition, the abun- However, for the ECMWF ensemble mean and control dant moisture transported to the YHRV is predicted run, the forecast of precipitation accumulation for event well by ECMWF at the lead time of 6 days, as the ver- 2 shows an increasing magnitude in northward deviation ifications show in Figs. 7c1 and 7c2. Since the conditions with increasing lead time. This demonstrates that the of long duration and abundant moisture are depicted meridional displacement of the PEPE from ECMWF is well in the forecasts from ECMWF, attention is then not completely dominated by the forecast deviations of turned to vertical velocity. Strong ascending motion, the key large-scale systems. Therefore, the reasonable which provides a mechanism for triggering convection forecast errors for the key large-scale systems are unable and large-scale moisture convergence, is also a signifi- to fully account for the deficiency of the ECMWF in cant condition for heavy rainfall, with its importance PEPE forecasts, especially for the meridional location of stressed in many studies (Lau and Kim 2012; Lenderink the rainband. For this reason, further investigation is et al. 2017). In view of the above facts, the description of conducted on the essential physical conditions that af- ascending motion in ECMWF is suspected to be a factor fect the production of extreme precipitation forecasts in the deficiency of ECMWF in the PEPE forecast. for KISAM and ECMWF. Therefore, efforts are devoted to the depiction of ver- From a rather macroscopic perspective, long duration, tical velocity at 500 hPa, which affects the production of abundant moisture, and strong ascending motion are precipitation forecast in ECMWF. For comparison, the vital physical conditions for extreme heavy pre- vertical velocity at 500 hPa related to the precipitation cipitation, on top of the conditions needed for ordinary production process in KISAM is also investigated.

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC 234 WEATHER AND FORECASTING VOLUME 33

FIG. 9. Observed and predicted time–latitude cross sections of vertical velocity at 500 hPa during event 1: (a1)–(a3) ECMWF ensemble mean forecasts; (b1)–(b3) ECMWF control forecast initialized at 12, 15, and 17 Jun; (c1)–(c3) vertical velocity at 500 hPa averaged over the three most similar historical analogous days that KISAM finds for forecasts initialized at 9, 12, and 15 Jun; and (d) the observed time– latitude cross section of the vertical velocity during event 1. The initializing days are marked in the bottom-left corners.

ECMWF provides a direct product of the vertical of the rainband during this event. Predictions of the velocity forecast in the medium forecast range of up to vertical velocity from the ECMWF ensemble mean 10-day lead. However, KISAM does not use the ver- with a forecast step of 6 h basically reproduce the shift in tical velocity forecast from NWP models, nor does it the observations at the lead times of 1 and 3 days directly produce forecasts of vertical velocity. Since (Figs. 9a2 and 9a3). However, the ECMWF ensemble KISAM is based on an analog method, its pre- mean overpredicts the intensity of the ascending motion cipitation forecast is actually regulated by the corre- at all initialization dates. Furthermore, the forecasts of sponding weather conditions on the analogous historical the time of occurrence of the strong ascending motion days. Hence, vertical velocity, which has an impact on the from the ECMWF ensemble mean are not quite con- production of precipitation forecasts in KISAM, is ob- sistent with the observations. When the forecast is ini- tained from the analogous days that KISAM finds at dif- tialized on 12 June for event 1, the ECMWF ensemble ferent lead times. mean fails to capture the northward shift of the 500-hPa Vertical velocities at 500 hPa that affect the pre- vertical velocity (Fig. 9a1). The forecasts of vertical cipitation output of the ECMWF ensemble mean, the velocity from the ECMWF ensemble mean deviate no- ECMWF control forecast, and KISAM are verified. The tably to the north of the observations. Interestingly, the observed meridional evolution of vertical velocity at performance of the ECMWF ensemble mean in pre- 500 hPa during event 1 is illustrated as a time–latitude dicting the location of the vertical velocity corresponds cross section in Fig. 9d. The vertical velocity also shifts well to its performance in capturing the location of from south to north in parallel with the northward shift precipitation at different lead times during event 1. The

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC FEBRUARY 2018 Z H O U E T A L . 235

FIG. 10. As in Fig. 9, but for event 2. directions of the meridional deviation for the forecasts times (Figs. 10a1–a3). The degree of the northward de- of vertical velocity and the accumulated precipitation viation becomes more severe as the lead time increases, from the ECMWF ensemble mean are consistent as lead which corresponds to the larger northward deviation of time varies. Forecasts of vertical velocity from the precipitation accumulation with increasing lead time. ECMWF control run at all initialization times are The forecast deviations of the ECMWF control run for stronger and more discontinuous in time. The depiction vertical velocity show similar trends to the meridional of the meridional evolution of the vertical velocity displacement of the whole precipitation process is similar to that of the ECMWF ensemble mean (Figs. 10b1–b3). With the exception of the first 2 days, (Figs. 9b1–b3). Vertical velocities from KISAM fail to which KISAM fails to predict in event 2, the vertical show the northward shift at all lead times (Figs. 9c1–c3). velocities related to precipitation production in KISAM Nevertheless, the vertical velocity from KISAM is lo- show an apparent advantage over those depicted in cated between 268 and 348N, with close agreement to the ECMWF, at longer lead times (Figs. 10c1–c3). There- observations. This means the evolution of vertical ve- fore, the quality of the depictions of the vertical velocity locity from KISAM coincides to a large extent with the during a PEPE for ECMWF and KISAM has a domi- observed meridional evolution. For the forecast initial- nant influence on their performance in the forecast of a ized on 9 June, due to the northward deviation of ver- PEPE at various lead times. The northward deviation of tical velocity in the ECMWF ensemble mean, KISAM the ascending motion is the leading cause of the de- shows better performance in predicting the vertical ve- ficiency of ECMWF in forecasting the meridional loca- locities in the process of precipitation production. For tion of the PEPE at longer lead times. In contrast, event 2, similar results are seen. As depicted in Fig. 10, vertical velocity related to precipitation production in predictions of vertical velocity from the ECMWF en- KISAM is achieved from the NCEP–NCAR reanalysis, semble mean show a clear northward bias at all lead which appears to be better during PEPEs compared to

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC 236 WEATHER AND FORECASTING VOLUME 33 the ECMWF forecast. This provides a proper elucida- the meridional deviation in the ECMWF ensemble mean tion for the superiority of KISAM for the forecast of forecasts of vertical velocity and accumulated pre- PEPEs at longer lead times. cipitation show good correspondence with varying lead time. In addition, the vertical velocities obtained from the NCEP–NCAR reanalysis related to precipitation pro- 6. Conclusions and discussion duction in KISAM show an apparent advantage over Two PEPEs during summer 2016 that caused severe those depicted in ECMWF, at longer lead times. There- flooding in the YHRV posed a serious challenge to op- fore, the quality in the depiction of vertical velocity dur- erational forecasters. In accordance with the regional ing PEPE for ECMWF and KISAM has a dominant PEPE definition introduced by Chen and Zhai (2013), influence on their performance in the forecasting of the these two cases are selected as verification cases for two meridional location of the PEPE at various lead times. objective forecast models. One is ECMWF, an out- For ECMWF, a hydrostatic model, the vertical ve- standing NWP model from TIGGE, and the other is a locity is determined by horizontal advection and pres- statistical downscaling model using the large-scale at- sure gradients, which results in inaccuracy. Moreover, mospheric circulation output from ECMWF, called errors in other complicated parameterization schemes, KISAM. An assessment and comparison of the two and initial and boundary conditions, may also con- forecast models is conducted for the two PEPEs, and the tribute to the inaccuracy of the precipitation forecast. factors affecting their forecast performance are ana- In contrast, statistical downscaling techniques like lyzed. The main conclusions that we draw are as follows. analog methods can easily establish nonlinear relation- For the location of daily heavy precipitation with ships between large-scale variables and local variables 2 amounts no less than 25 mm day 1 during the PEPE, (Fernández and Sáenz 2003). For KISAM, the vertical about one-half or more of the possible solutions in the velocity that actually regulates the production of the spectrum of the ECMWF ensemble show good skill at precipitation forecast is the average vertical velocity all lead times. However, the ECMWF ensemble shows a obtained from the NCEP–NCAR reanalysis of the three reduction in skill as the lead time increases, and it most analogous historical records. Therefore, the im- shows a general underestimation of heavy precipitation proved capture of the ascending motion related to pre- 2 ($25 mm day 1). At lead times longer than 3 days, cipitation production in KISAM explains its superior KISAM outperforms nearly one-half or more of all the performance compared to the ensemble mean and half ensemble members and the ensemble mean of ECMWF, the members of ECMWF for the PEPE at longer lead especially for event 1. Moreover, at longer lead times, times. In addition, based on the very thorough un- KISAM also shows an advantage over the ECMWF derstanding of the circulation patterns conducive to ensemble and the control forecasts in terms of the me- PEPEs as described by Chen and Zhai (2014), KISAM ridional location of the two PEPE processes. It is in- combines the key circulation variables from the typical teresting that KISAM achieves higher ETSs at longer circulation patterns together as predictors. This may lead times, which is also found for the ECMWF control help KISAM search corresponding analogous patterns forecast. This can be attributed to the fact that the in the PEPE historical pool more precisely and produce evaluation is conducted against only two PEPE cases, better forecasts of PEPEs. while the general tendency of decreasing skill as lead In summary, the ECMWF ensemble provides a skill- time increases is based on a larger number of samples. ful envelope of solutions for PEPE forecasts. However, Since the large-scale variables from the ECMWF en- as with other NWP models, the ECMWF forecast per- semble mean forecast are used by KISAM as predictors, formance declines as lead time increases. In view of this, the performances of KISAM and ECMWF for PEPE statistical downscaling techniques can serve as an im- forecasts are both related to the forecast performance of portant aid for forecasters to conduct forecasts of heavy these key large-scale systems during PEPE. In view of precipitation in a medium forecast range of 4–10 days. this, the ability of the ECMWF ensemble mean in pre- Moreover, the forecast verification scores of the ECMWF dicting key large-scale systems during the PEPE is ver- ensemble and KISAM at longer times are still low. There ified for lead times of up to 9 days. This demonstrates is still a great need to improve our predictions of PEPEs, that good correspondence between the deviation in whether with statistical downscaling models or perhaps latitude of the key large-scale systems and the meridi- with a high-resolution dynamic model with state-of-the- onal displacement of the precipitation accumulation art data assimilation. does not exist. According to further verifications of the ascending motion, which affects the production of heavy Acknowledgments. This study was jointly supported precipitation in ECMWF and KISAM, the directions of by the National Key Technology Support Program

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC FEBRUARY 2018 Z H O U E T A L . 237

(Grant 2015BAC03B02) and the Basic Research Fund Karl, T., and R. Knight, 1998: Secular trends of precipitation of the Chinese Academy of Meteorological Sciences amount, frequency, and intensity in the United States. (2015Z001). The authors are grateful to the editor Brian Bull. Amer. Meteor. Soc., 79, 231–241, https://doi.org/10.1175/ 1520-0477(1998)079,0231:STOPAF.2.0.CO;2. Ancell and three anonymous reviewers for their in- Kobold, M., and K. Suselj, 2005: Precipitation forecasts and their valuable and constructive suggestions and comments uncertainty as input into hydrological models. Hydrol. Earth that helped improve the manuscript. Syst. Sci., 9, 322–332, https://doi.org/10.5194/hess-9-322-2005. Krishnamurti, T. N., A. D. Sagadevan, A. Chakraborty, A. K. Mishra, and A. Simon, 2009: Improving multimodel weather REFERENCES forecast of monsoon rain over China using FSU super- ensemble. Adv. Atmos. Sci., 26, 813–839, https://doi.org/ Ben Daoud, A., E. Sauquet, G. Bontron, C. Obled, and M. Lang, 10.1007/s00376-009-8162-z. 2016: Daily quantitative precipitation forecasts based on the Lau, W. K., and K. Kim, 2012: The 2010 Pakistan flood and Russian analogue method: Improvements and application to a French heat wave: Teleconnection of hydrometeorological ex- large river basin. Atmos. Res., 169, 147–159, https://doi.org/ tremes. J. Hydrometeor., 13, 392–403, https://doi.org/10.1175/ 10.1016/j.atmosres.2015.09.015. JHM-D-11-016.1. Caillouet, L., J. P. Vidal, E. Sauquet, and B. Graff, 2016: Proba- Lavers, D. A., and G. Villarini, 2013: Were global numerical bilistic precipitation and temperature downscaling of the weather prediction systems capable of forecasting the extreme Twentieth Century Reanalysis over France. Climate Past, 12, Colorado rainfall of 9–16 September 2013? Geophys. Res. 635–662, https://doi.org/10.5194/cp-12-635-2016. Lett., 40, 6405–6410, https://doi.org/10.1002/2013GL058282. Chen, Y., and P. M. Zhai, 2013: Persistent extreme precipitation Lenderink, G., R. Barbero, J. M. Loriaux, and H. J. Fowler, 2017: events in China during 1951–2010. Climate Res., 57, 143–155, Super-Clausius–Clapeyron scaling of extreme hourly convec- https://doi.org/10.3354/cr01171. tive precipitation and its relation to large-scale atmospheric ——, and ——, 2014: Two types of typical circulation pattern for conditions. J. Climate, 30, 6037–6052, https://doi.org/10.1175/ persistent extreme precipitation in central–eastern China. JCLI-D-16-0808.1. Quart. J. Roy. Meteor. Soc., 140, 1467–1478, https://doi.org/ Lu, R. Y., 2000: Anomalies in the tropics associated with the heavy 10.1002/qj.2231. rainfall in East Asia during the summer of 1998. Adv. Atmos. Dee, D., and Coauthors, 2011: The ERA-Interim reanalysis: Sci., 17, 205–220, https://doi.org/10.1007/s00376-000-0004-y. Configuration and performance of the data assimilation sys- Niu, R. Y., and P. M. Zhai, 2013: Synoptic verification of medium- tem. Quart. J. Roy. Meteor. Soc., 137, 553–597, https://doi.org/ extended-range forecasts of the northwest Pacific subtropical 10.1002/qj.828. high and South Asian high based on multi-center TIGGE Ding, Y. H., 1992: Summer monsoon rainfalls in China. J. Meteor. data. Acta Meteor. Sin., 27, 725–741, https://doi.org/10.1007/ Soc. Japan, 70, 373–396, https://doi.org/10.2151/ s13351-013-0513-0. jmsj1965.70.1B_373. ——, ——, and B. Q. Zhou, 2015: Evaluation of forecast perfor- ——, and E. Reiter, 1982: A relationship between planetary waves mance of Asian summer monsoon low-level winds using the and persistent rain and thunderstorms in China. Theor. Appl. TIGGE dataset. Wea. Forecasting, 30, 455–470, https://doi.org/ Climatol., 31, 221–252. 10.1175/WAF-D-13-00141.1. Ebert, E. E., U. Damrath, W. Wergen, and M. E. Baldwin, 2003: Park, Y. Y., R. Buizza, and M. Leutbecher, 2008: TIGGE: Pre- The WGNE assessment of short-term quantitative pre- liminary results on comparing and combining ensembles. cipitation forecasts. Bull. Amer. Meteor. Soc., 84, 481–492, Quart. J. Roy. Meteor. Soc., 134, 2029–2050, https://doi.org/ https://doi.org/10.1175/BAMS-84-4-481. 10.1002/qj.334. Fan, K., H. J. Wang, and Y. J. Choi, 2008: A physically-based sta- Pelly, J. L., and B. J. Hoskins, 2003: How well does the ECMWF Ensemble Prediction System predict blocking? Quart. J. Roy. tistical forecast model for the middle-lower reaches of the Meteor. Soc., 129, 1683–1702, https://doi.org/10.1256/qj.01.173. Yangtze River valley summer rainfall. Chin. Sci. Bull., 53, 602– Ralph, F. M., and Coauthors, 2005: Improving short-term (0–48 h) 609, https://doi.org/10.1007/s11434-008-0083-1. cool season quantitative precipitation forecasting: Recommen- Fernández, J., and J. Sáenz, 2003: Improved field reconstruction dations from a USWRP workshop. Bull. Amer. Meteor. Soc., 86, with the analog method: Searching the CCA space. Climate 1619–1632, https://doi.org/10.1175/BAMS-86-11-1619. Res. , 24, 199–213, https://doi.org/10.3354/cr024199. ——, E. Sukovich, D. Reynolds, M. Dettinger, S. Weagle, Galarneau, T. J., T. M. Hamill, R. M. Dole, and J. Perlwitz, 2012: A W. Clark, and P. J. Neiman, 2010: Assessment of extreme multiscale analysis of the extreme weather events over quantitative precipitation forecasts and development of re- western Russia and northern Pakistan during July gional extreme event thresholds using data from HMT-2006 2010. Mon. Wea. Rev., 140, 1639–1664, https://doi.org/10.1175/ and COOP observers. J. Hydrometeor., 11, 1286–1304, https:// MWR-D-11-00191.1. doi.org/10.1175/2010JHM1232.1. Hartmann, D. L., and Coauthors, 2013: Observations: Atmosphere Samel, A. N., W. C. Wang, and X. Z. Liang, 1999: The monsoon and surface. Climate Change 2013: The Physical Science Basis, rainband over China and relationships with the Eurasian T. F. Stocker et al., Eds., Cambridge University Press, 159–254. circulation. J. Climate, 12, 115–131, https://doi.org/10.1175/ Hirsch, R. M., and S. A. Archfield, 2015: Flood trends: Not higher 1520-0442-12.1.115. but more often. Nat. Climate Change, 5, 198–199, https://doi.org/ Sharma, S., and Coauthors, 2017: Eastern U.S. verification of en- 10.1038/nclimate2551. semble precipitation forecasts. Wea. Forecasting, 32, 117–139, Kalnay, E., and Coauthors, 1996: The NCEP/NCAR 40-Year Re- https://doi.org/10.1175/WAF-D-16-0094.1. analysis Project. Bull. Amer. Meteor. Soc., 77, 437–471, https:// Su, X., H. Yuan, Y. Zhu, Y. Luo, and Y. Wang, 2014: Evaluation doi.org/10.1175/1520-0477(1996)077,0437:TNYRP.2.0.CO;2. of TIGGE ensemble predictions of Northern Hemisphere

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC 238 WEATHER AND FORECASTING VOLUME 33

summer precipitation during 2008–2012. J. Geophys. Res. At- Atmos. Res., 103, 106–118, https://doi.org/10.1016/ mos., 119, 7292–7310, https://doi.org/10.1002/2014JD021733. j.atmosres.2011.04.003. Sukovich, E. M., F. M. Ralph, F. E. Barthold, D. W. Reynolds, and Zhai, P., X. Zhang, H. Wan, and X. Pan, 2005: Trends in total D. R. Novak, 2014: Extreme quantitative precipitation fore- precipitation and frequency of daily precipitation extremes cast performance at the Weather Prediction Center from 2001 over China. J. Climate, 18, 1096–1108, https://doi.org/10.1175/ to 2011. Wea. Forecasting, 29, 894–911, https://doi.org/10.1175/ JCLI-3318.1. WAF-D-13-00061.1. Zhi, X. F., X. D. Ji, J. Zhang, L. Zhang, Y. Q. Bai, and C. Z. Lin, Swinbank, R., and Coauthors, 2016: The TIGGE project and its 2013: Multimodel ensemble forecasts of surface air tempera- achievements. Bull. Amer. Meteor. Soc., 97, 49–67, https://doi.org/ ture and precipitation using TIGGE datasets (in Chinese). 10.1175/BAMS-D-13-00191.1. Trans. Atmos. Sci., 36, 257–266. Tao, S. Y., 1980: The Torrential Rain in China. Science Press, 225 pp. Zhou, B. Q., and P. M. Zhai, 2016: A new forecast model based Wang, C., 2014: On the calculation and correction of equitable threat on the analog method for persistent extreme precipita- score for model quantitative precipitation forecasts for small tion. Wea. Forecasting, 31, 1325–1341, https://doi.org/10.1175/ verification areas: The example of Taiwan. Wea. Forecasting, 29, WAF-D-15-0174.1. 788–798, https://doi.org/10.1175/WAF-D-13-00087.1. ——, R. Y. Niu, and P. M. Zhai, 2015: An assessment of the pre- Wang, H. J., and K. Fan, 2009: A new scheme for improving dictability of the East Asian subtropical westerly jet based on the seasonal prediction of summer precipitation anomalies. TIGGE data. Adv. Atmos. Sci., 32, 401–412, https://doi.org/ Wea. Forecasting, 24, 548–554, https://doi.org/10.1175/ 10.1007/s00376-014-4026-2. 2008WAF2222171.1. Zhu, L., and Coauthors, 2010: Characteristics of climate change Wilby, R. L., and T. M. L. Wigley, 1997: Downscaling general of south Asia high in summer and its impact on pre- circulation model output: A review of methods and limita- cipitation in eastern China (in Chinese). Plateau Meteor., tions. Prog. Phys. Geogr., 21, 530–548, https://doi.org/10.1177/ 29, 671–679. 030913339702100403. Zorita, E, and H. Von Storch, 1999: The analog method as a simple Willems, P., K. Arnbjerg-Nielsen, J. Olsson, and V. T. V. Nguyen, statistical downscaling technique: Comparison with more 2012: Climate change impact assessment on urban rainfall complicated methods. J. Climate, 12, 2474–2489, https://doi.org/ extremes and urban drainage: Methods and shortcomings. 10.1175/1520-0442(1999)012,2474:TAMAAS.2.0.CO;2.

Unauthenticated | Downloaded 10/04/21 05:24 PM UTC