Improvement and Ensemble Strategy of Heavy-Rainfall Quantitative Precipitation Forecasts using a Cloud-Resolving Model in Taiwan

Transcription

1 Improvement and Ensemble Strategy of Heavy-Rainfall Quantitative Precipitation Forecasts using a Cloud-Resolving Model in Taiwan Chung-Chieh Wang Department of Earth Sciences, National Taiwan Normal University, Taipei, Taiwan Acknowledgements: SH Chen, BY Chuang, CS Chang, SY Huang, CC Huang, YW Wang, TTFRI, NCHC Workshop on Disaster Risk Reduction in Extreme Climate 23 November 2017, Taipei, Taiwan

2 Presentation outline 1. Why Cloud-Resolving Models for QPFs? 1.1 Real-Time Cloud-Resolving Forecasts by CReSS in Taiwan 1.2 Verification of 24-h QPFs: Standard Categorical Matrices 1.3 Current Skill Level for Typhoon QPFs 2. How CReSS Performs at Short Range (0-72 h)? 2.1 For 24-h typhoon QPFs: (Six) Seasons 2.2 For 24-h Mei-yu QPFs: (3 Seasons) 3. How to Get Ensemble and Improve Further? 3.1 A Strategy to Obtain Cloud-Resolving QPFs at Longer Range 3.2 Usefulness of Daily Time-Lagged Ensemble in Six-hourly Time-Lagged Ensemble for Three Typhoons 4. Conclusion and Summary 2

3 1. Why Cloud-Resolving Models for QPFs? 1.1 Real-Time Cloud-Resolving Forecasts by CReSS in Taiwan Real-time CReSS forecasts since , non-stop year-round since 2010 Gradual increases in resolution, forecast length, and domain size Using NCEP GFS 1 x 1 analyses/forecasts as IC/BCs (0.5 x 0.5 since 2013) Current forecasts (40L) every 6 h out to 78 h: 5 km (216 x 180) 2.5 km (600 x 480) NTNU/Department of Earth Sciences CReSS 2.5km Realtime Forecast Routinely provided to domain in TTFRI of Taiwan as the only cloud-resolving member ( x = 2.5 km) Real-time (and all past) results available at (Current 2.5-km domain: 1500 km x 1200 km) 3

4 Real-time forecast website at 5-, 3.5-, 4-, and 5-km forecasts through the years since km 3-day forecasts since 2010 (currently 1500 km x 1200 km) 2.5-km 8-day forecasts (currently 1860 km x 1360 km) 4

5 1.2 Verification of 24-h QPFs: Standard Categorical Matrices Commonly-used verification methods for QPFs: Both subjective (visual) and objective verifications Widely-used skill scores: Threat score (TS), bias score (BS), probability of detection (POD), false-alarm rate (FAR), and odds ratio (OR) 24-h accumulative rainfall (day 1, 2, 3), from forecasts starting at 0000 or 1200 UTC Rainfall thresholds: 0.05 to 1000 mm Evaluated on rain-gaugesites (about 450 points) with equal weight OBS M N CN H FA FCST TS = H / (M + H + FA) BS = (H + FA) / (H + M) POD = H / (H + M) FAR = FA / (H + FA) 5

6 1.3 Current skill level for typhoon QPFs Current skill level of QPFs for typhoons: Comparison of 2014 between Global vs. regional models (left): Regional models win! Regional mesoscale vs. CReSS at TTFRI (right): The CRM wins! At CWB and major centers At TTFRI, for two TCs in 2014 TAPEX-26, WRF-20, MM5-2, CReSS h (day 1) QPF Official Global Regional Huang et al. (2015): CWB conference 6-30 h QPF (mm) TS ~ 0.2 at 300 mm, little skill 500 mm (per 24 h) for day-1 QPFs Threshold (mm, per 24 h) JJ Wang et al. (2015): CWB conference 6

7 2. How CReSS Performs at Short Range (0-72 h)? 2.1 For 24-h typhoon QPFs: (Six) Seasons Classify h segments (warning periods) based on observed rainfall Results from all segments (no classification): All Sub-groups with peak 24-h rainfall reaching 200 mm (98), 350 mm (52), 500 mm (26), and 750 mm (14, roughly top 7%) Real-time results of Morakot (2009) with peak amount ~ 1560 mm 7

8 TS: Classification based on observed 24-h peak rainfall Day Least rain Most rain 0.15 Day 2 Day Updated from Wang et al. (2013, J Hydrol), Wang (2015, MWR), Wang (2016, BAMS) 8

9 BS: Weak tendency of under-forecast at high thresholds toward day 3 Day Total points at each threshold: Color number: observed pts (i.e., O ) Least rain 3 3 Most rain Day 2 Day 3 9

10 An example of real-time 2.5-km forecasts made at 7/10 12 UTC, for Soulik (2013), from 7/12 00 UTC to 7/13 18 UTC (t = h): Soulik (2013) Model forecasts of MSLP, sfc winds, and 1-h rainfall (mm) CReSS initial time Radar reflectivity composite (dbz) 10

11 Real-time 2.5-km forecasts made at 7/10 12 UTC, for Soulik (2013): OBS 876 mm 7/ Z 7/ Z 7/ Z OBS M H FA FCST CReSS TS = H / (M + H + FA) BS = (H + FA) / (H + M) Day 1 Day 2 Day 3 Day 1 Day 2 Day 3 Threshold (mm) Threshold (mm) 11

12 Other examples of high-resolution deterministic forecasts by CReSS: Morakot (2009): Most devastating typhoon, extreme rainfall up to 2855 mm Real-time twice daily 4-km CReSS forecasts (for 48 h) in 2009 Total 48-h rainfall (mm) from the run starting at 12 UTC 6 Aug 2009: Day-2 forecast from the run starting at 00 UTC 7 Aug 2009: OBS CReSS OBS CReSS 8/ Z 8/6 12Z-8/8 12Z (Wang et al. 2013: J. Hydro.) TS BS (Wang 2014: WAF) 24-h rainfall threshold (mm) 12

13 2.2 For 24-h Mei-yu QPFs: (3 Seasons) Total of 337 segments in Mei-yu period (May-Jun, excluding TC periods), classified into: Groups A, B, C, D, and X from the most to the least overall amount At least 10% of sites 50 mm (A), 25 mm (B), 10 mm (C), 1 mm (D), or otherwise (X, i.e., almost no rainfall) Group A+: 10% sites 130 mm, all from group A (roughly top 4% of all sample, with highest hazard potential) Total sample size: 337 segments ( points) May-Jun (excluding TCs) D X A A+: 13 (~4%) A group: 61 (27143) B group: 75 (33285) Compute scores after summing entries from all segments into one 2 x 2 contingency table C B C group: 88 (33846) D group: 67 (29443) X group: 46 (25250)

14 TS: Day 1 (0-24 h) Most rain Least rain Day 2 (24-48 h) Day 3 (48-72 h)

15 3. How to Get Ensemble and Improve Further? 3.1 A Strategy to Obtain Cloud-Resolving QPFs at Longer Range Time-lagged ensemble (with comparable cost to multi-member ensemble): Wang et al. (2016, WAF), Wang et al. (2017, BAMS) 15

16 3.2 Usefulness of Daily Time-Lagged Ensemble in Domain size: 1860 km x 1360 km (744 x 544 pts), forecast range: 8 days Results from daily forecasts (t 0 = 00 UTC) for 6 TCs in Talim (TY1205) Doksuri (TY1206) Saola (TY1209) Tembin (TY1214) 8-day fcst domain Jelawat (TY1217) Kong-Rey (TY1315) 16

17 Forecast for Typhoon Saola, starting at 00 UTC 30 Jul 2012 CWB best-track CReSS initial time TTFRI ensemble at 7/30 00 UTC (Computational demand: ~5 TTFRI-WRF members) (~10 km/h) (~24 members) Ensemble tracks from time-lagged runs 17

18 Comparison with TRMM-PR rain-rates for realistic scenarios: Same forecast starting at 00 UTC 30 Jul / UTC 7/ UTC 8/ UTC t = 29 h t = 45 h t = 52 h 7/ UTC 7/ UTC 8/ UTC 18

19 QPFs targeted at the most-rainy date, one from each of five typhoons: For Talim (6/20), Doksuri (6/28), and Saola (8/2): D8 D4 D6 Black, red, blue, and green box: TS 0.2 at 50, 100, 200, and 350 mm 19

20 QPFs targeted at the most-rainy date, one from each of five typhoons: For Tembin (8/24), Jelawat (9/28), and Kong-Rey (8/29): D3 D6 D7 Black, red, blue, and green box: TS 0.2 at 50, 100, 200, and 350 mm 20

21 Threat score TS of a decent QPF made at the longest range for the most rainy date of each typhoon (from once daily QPFs): A decent QPF at a range beyond day-3 expect for Tembin Averaged range of such decent QPFs is on day 5.7! Day 7 Day 3 Day 6 Day 6 Day 4 Day 8 Rainfall threshold (mm, per 24 h) Observed peak rainfall Wang et al. (2016): WAF, Wang et al. (2017): BAMS 21

22 3.3 Six-hourly Time-Lagged Ensemble for Three Typhoons First example: TS Kong-Rey (2013) Very weak and not expected to cause serious damages Most hazardous TY in 2013 (6 deaths and NTD 900 M in damage) Forecast for Kong-Rey, starting at 06 UTC 22 Aug 2013 CWB best-track TTFRI Kong-Rey (TY1315) CReSS initial time Focus: Ability to provide all possible scenarios at longer lead time 22

23 Comparison of tracks with typical 24-member 3-day ensemble 00Z 22-00Z 25 Aug Current 24-member Ensemble at TTFRI First release at 00 UTC 27 Aug 2013 Black: OBS 00Z 25-18Z 28 Aug Tracks produced by timelagged ensemble: Before 00 UTC 25 Aug 2013: cover all possible scenarios After 00 UTC 25 Aug 2013: tracks converge toward the most-likely case Black: OBS Track (and QPF scenarios) available much earlier by time-lagged ensemble 23

24 The produced 48-h QPFs (29-30 Aug) for Taiwan: Highly realistic rainfall scenario in association with each track due to cloud-resolving capability About 2/3 members indicate heavy rainfall potential over plains First 12 members (before 8/26 00Z) Last 13 members (8/26-28) (48-h rainfall) Wang et al. (2016): WAF, Wang et al. (2017): BAMS 24

25 Probability (ensemble) information generated by the time-lagged ensemble: The evolving probabilities give forecasters a good sense of what to expect 100 mm 200 mm 350 mm 500 mm First 12 members (before 8/26 00Z) Last 13 members (8/26-28) 25

26 Second example: TY Soulik (2013): Second most hazardous TY in 2013 (after TS Kong-Rey) 3 deaths, NTD 1.5 B in damage Soulik (2013) 7/ Z 7/ Z TTFRI 876 mm 8-day fcst domain CReSS initial time 7/ Z 7/ Z 26

27 Comparison of tracks with typical 24-member 3-day ensemble Current 24-member Ensemble at TTFRI First release at 18 UTC 9 Jul 2013 Soulik (2013) 7/6 00 UTC to 7/10 12 UTC Black: OBS 7/13 00 Z Tracks produced by timelagged ensemble: Before 12 UTC 10 Jul 2013: comparable spread but more to the north After 12 UTC 10 Jul 2013: tracks converge toward BT 7/6 00 UTC to 7/12 18 UTC Black: OBS 7/13 00 Z Larger cross-track errors (in moving direction) prior to 10 July 27

28 The produced 24-h QPFs (7/ UTC) for Taiwan: 27 members total 7/6 00Z 7/6 06Z 7/6 12Z 7/6 18Z 7/7 00Z 7/7 06Z 7/7 12Z 7/7 18Z 7/8 00Z 7/8 06Z 7/8 12Z 7/8 18Z OBS 7/ Z 876 mm 7/9 00Z 7/9 06Z 7/9 12Z 7/9 18Z 7/10 00Z 7/10 06Z 7/10 12Z (24-h rainfall) 7/10 18Z 7/11 00Z 7/11 06Z 7/11 12Z 7/11 18Z 7/12 00Z 7/12 06Z 7/12 12Z 28

29 Probability (ensemble) information generated for Soulik: (black = OBS) 50 mm 100 mm 200 mm 300 mm 500 mm First 9 members before 7/8 06Z Middle 9 mem., 7/8 06Z- 7/10 06Z Last 9 members after 7/10 06Z 29

30 Third example: TY Soudelor (2015) Most hazardous TY in deaths, 4 missing, NTD 2.2 B in damage, 4 M houses no power Soudelor (2015) 8/ UTC 8/ UTC TTFRI 8-day fcst domain CReSS initial time 8/ UTC 8/ UTC 30

31 CReSS forecast made at 8/4 06 UTC for Soudelor, loop from 8/7 06 UTC to 8/8 12 UTC (t = h): Soudelor (2015) Model forecasts of MSLP, sfc winds, and 1-h rainfall (mm) CReSS initial time Radar reflectivity comp. (dbz) 8/7 06 UTC to 8/8 12 UTC 31

32 Comparison with TRMM-TMI T B (85v, 89v, 91v) for realistic TC structure: Same forecast starting at 06 UTC 4 Aug / UTC 8/ UTC 8/ UTC t = 60 h t = 76 h t = 94 h 8/ UTC 8/ UTC 8/ UTC 32

33 24-h QPFs and their comparison with the observation: OBS 8/ UTC 8/ UTC 8/ UTC CReSS t = h t = h t = h 33

34 Comparison of tracks with typical 24-member 3-day ensemble Current 24-member Ensemble at TTFRI First release at 00 UTC 5 Aug 2015 Soudelor (2013) 8/1 00 UTC to 8/4 00 UTC Black: OBS 8/7 00 Z Tracks produced by timelagged ensemble: Before 00 UTC 4 Aug 2015: cover all possible scenarios After 00 UTC 4 Aug 2015: tracks converge toward the most-likely case 8/1 00 UTC to 8/7 18 UTC 8/7 00 Z Black: OBS Early tracks (1-2 Aug) had speed (along-track) errors, too fast by 6-12 h 34

35 The produced 24-h QPFs (7/ UTC) for Taiwan: 29 members total 7/31 12Z 7/31 18Z 8/1 00Z 8/1 06Z 8/1 12Z 8/1 18Z 8/2 00Z OBS 8/7-8 12Z 8/2 06Z 8/2 12Z 8/2 18Z 8/3 00Z 8/3 06Z 8/3 12Z 8/3 18Z 8/4 00Z 8/4 06Z 8/4 12Z 8/4 18Z 8/5 00Z 8/5 06Z 8/5 12Z (24-h rainfall) 8/5 18Z 8/6 00Z 8/6 06Z 8/6 12Z 8/6 18Z 8/7 00Z 8/7 06Z 8/7 12Z 35

36 Probability (ensemble) information generated for Soudelor: (black = OBS) 50 mm 100 mm 200 mm 300 mm 500 mm First 10 members before 8/3 00Z Middle 10 mem., 8/3 00Z- 8/5 06Z Last 9 members after 8/5 06Z 36

37 Operational 2.5-km CReSS shows high skills in heavy-rainfall QPFs in Taiwan for both typhoon and mei-yu, not limited to 0-24 h but also in days 2-3 Significantly improved skill using CReSS at cloud-resolving resolution Better skill for top events than all events, not understood previously For mei-yu QPFs, TS > 0.15) through 500/200/130 mm on day 3/2/1 ~ Season Threshold 50 mm 130 mm 200 mm 350 mm 500 mm Typhoon Top 7% cases ( ) Mei-yu Top 4% cases ( ) 4. Conclusion and Summary Day 1 (0-24 h) Day 2 (24-48 h) Day 3 (48-72 h) Day 1 (0-24 h) Day 2 (24-48 h) Day 3 (48-72 h) Much higher predictability over terrain for systems linked to topography (e.g., topo. uplift), where high skill at high thresholds can be achieved Time-lagged ensemble allows for high-resolution, large fine-domain (for QPFs), and longer range (for lead time and preparation) simultaneously

38 At longer range beyond 3-4 days: High forecast uncertainty and large spread in TC tracks Wide range of scenarios, including worse case for early preparation At shorter range as the TC approaches (within 3-4 days): TC tracks converge as forecast uncertainty reduces, and the most-likely scenario emerges (which is, with high chances, known) High probability of heavy rainfall, highly realistic in distribution due to strong topographic control Authority can then make adjustments for the most-likely scenario 38

39 --- The End --- Thanks for your for attention! Questions?