Extended Ensemble Nowcasting Technique Using Recognition on

Toward the Extended Ensemble
Nowcasting Technique Using Pattern
Recognition on Radar and WEPS
Trengshi Huang PhD.
Weather Forecast Center, Central Weather Bureau
Multi-stage Quantify Precipitation Forecast
Weekly (Qualitative)
GFS, Synaptic Analysis, statistics, analogy,
conceptual
Traditional 3d/4d
data assimulatoin
Storm scale data
assimulation
Radar
Extrapolation
1-3 daily QPF (Quantitative)
Regional (Ensemble) Forecast System,Statistics,
anvanced Ensemble Forecast
0-6 hr (or 0-12 hr) QPF
LAPS/STMAS, ARPS, VDRAS, Cloud model
0-1 hr QPF
Pattern
Recognition ?
Radar extrap., ANC, SCAN
0 hr Nowcasting
Radar, gagues,
lighting
2
Very Short Term Forecast/Nowcast
Monitor and Nowcasting Systems by Weather Satellite Center
測試循環式(cycling)雷
達資料同化策略中
(W.P. Huang, WSC)
Big data and Ensemble Forecast
 Big data is an all-encompassing term for any
collection of data sets so large and complex
that it becomes difficult to process using
traditional data processing applications.
 The challenges include analysis, capture,
curation, search, sharing, storage, transfer,
visualization, and privacy violations.
 如何在大量的系集預報資料中取得有用的
資訊？
http://en.wikipedia.org/wiki/Big_data
Basic Ensemble QPF Products
 Most probable single solution (deterministic forecast ):
Ensemble mean, median, etc…
QPESUMS Analysis
 CWB WRF EPS (Ensemble Prediction System): 20 members,
5km (李志昕&洪景山，2011)
 Take the ensemble mean among the ensemble dimension
 Forecast Variance (spread):
standard deviation, max., min., max. 10% mean, …
 Evaluation from mean and standard deviation
(丘台光、陳嘉榮、張保亮、林品芳，2007)
Mean
MAX.
STD
MIN
Advanced Ensemble QPF Products
 The ensemble average "smears" the rain rates so that the
maximum rainfall is reduced and area of light rain is enhanced
 Probability Density Function (PDF) approach on QPF
 The same spatial shape as the ensemble mean
 The same PDF of the entire ensemble system (PM, Elbert 2001).
 Or Averaging the PDFs among the ensemble dimension (newPM, 葉等2014)
 Deal not on spatial distribution.
Mean
PM
newPM
QPESUMS Analysis
From PQPFx to QPFP-y%
Select a threshold, say 50% chance, in probability space to make QPF
exceeding 50% chance
POP=PQPF0.1
Assign hatched
area to 0.1 mm
PQPF130
Assign hatched
area to 130 mm
PQPF10
Assign hatched
area to 10 mm
PQPF200
Assign hatched
area to 200 mm
PQPF25
Assign hatched
area to 25 mm
PQPF350
Assign hatched
area to 350 mm
PQPF50
Assign hatched
area to 50 mm
PQPF500
Assign hatched
area to 500 mm
PQPF100
QPFP-50%
Assign hatched
area to 100 mm
QPF exceeding y% probability threshold: QPFP-y%
Deterministic but probabilistic inside
 Easy for Forecasters to make decision
 Easy to use since it is QPF

QPFP-5%
MAX
QPFP-10%
QPFP-20%
QPFP-30%
QPFP-40%
QPFP-50%
MEAN
QPFP-60%
QPFP-70%
QPFP-90%
QPFP-100%
MIN
Predictability Evolution with Forecast Hours
1.0
Skill Score
 Radar QPF extrapolation
 High resolution NWPs &
Ensemble Prediction System
 Storm scale (radar) data
assimulation and forecast
NWPs
0 1
3
6
Radar QPF
Fct. hours
Integration on the Extended Nowcasting Precipitation
 Combining the CWB WRF-EPS (Li et al, 2013) and QPESUMS observation to
develop the extended nowcasting on QPF
 Objective: to improve the 0-6 h
QPF
 Key factors：
 Pattern recognigition
skill
(Moment invarant, Chen et al.
2014)
 Advanced Ensemble skill (Huang
et al, 2014; Yeh et al., 2014)
Integration on the Extended Nowcasting Precipitation
本研究
Originial
Radar based
(extrapolation)
結合外延及
模式校正
(ARMOR)
Combing the
observation.
WEPS
and
Radar
(NWP model)
Radar Data
Assimulation
…
1. Develop a pattern recognition to rank
the WEPS forecast during a time window.
2. Advanced Ensemble QPF
1.Pattern Recognition of Observed CV and WEPS-Based
QPF (PROCWB-QPF)
2.Radar-Ensemble Matching Algorithm (REMA)
Extended Nowcasting on QPF or QPN
(Chen et al., 2014)
Pattern Recognition (I)
 Moment invariants (Hu 1962)
OBS
WEPS
由動差不變量理論
得出具平移、尺度
和旋轉不變的七個
特徵描述值
 1O 
 
  
 O 

 7 
 1i 
 
 
 i 

 7
再由正規化相似度
演算法計算距離相
似度(DS)和角度相
似度(AS)
DS i  AS i
S 
2
i
 介於0到1之間，愈接
近 1表示模式與觀測
的相似程度愈高。
網格化的回
波值(dBZ)
(Chen et al., 2014)
Pattern Recognition (II)
 Piecewise recognition
S ni
S ni 1
S1i
S 2i
N
S
i
ave


S ni / N
n 1
(Chen et al., 2014)
Sampling, Recognizing, and Shifting
Radar CV
 Targeted time window (±6hr, 3hourly) based on
observation.
 22 members：WRFD, TWRF, WRF-EPS (20 members)
 4 Lag Runs
 5x22x4=440 samples
Fro ensemble member N (N=1, 2, …, 22)
forecast at target time (say 0 hr)
05/15 00Z
05/14 18Z
取4個
lag run
05/14 12Z
05/14 06Z
05/15 1200Z
-6
-3
6
9
12
15
18
21
24
18
21
24
27
30
24
27
30
33
36
‒6 h
0
3
6
1212h 15
18
+6 h
Shifting Forecast: Ranking 1 to 10 in 440 samples
Hindcast
Radar CV
05/15 1200Z
-3
0
3
6
9
12
Forecast
Rank
1-10
Ensemble and Advanced Ensemble Forecast
Hindcast
 Ensemble
Mean
among Ranks 1 to 10
 Ensemble
mean
smears
members’
extreme QPF values
 Advanced Ensemble
 Ensemble QPFP-20%
 Why 20%?
 At least 2 of 10
members
 Overestimate
the
overall
QPF
sometimes
 PM, newPM?
To be continue…
-3
0
3
6
9
12
Forecast
2014 May & Jun Verification on 0-3h QPF
 May and June verification
on 0-3h QPF.
 Observation Radar CV
larger then 10% for
recognition or not.
TS
POD
ETS
FAR
134/488 Cases, 37000(1902)/370000 km2
Predictability Evolution with Forecast Hours
1.0
Skill Score
 Radar QPF extrapolation
 High resolution NWPs &
Ensemble Prediction System
 Storm scale (radar) data
assimulation and forecast
 Pattern Recognition/ Single
Model Forecast/ Ensemble
Forecast
 Ensemble and Advanced
Ensemble Forecasting
 Fitting and Calibration??
Fitting?
NWPs
0 1
3
6
Ptn Rcg./EPS
Ptn Rcg./Single
Radar QPF
Fct. hours
Advanced Ensemble Nowcasting for Typhoon
Fung-Wang (2014) at 09/21 0000 UTC
CWB Operational Typhoon QPF
 Analog Approach
 Historical Typhoon tracks and precipitation data base
 Climatology Approach
 Climatological Typhoon positions and precipitation rate
 Numerical Weather Prediction
 Global Forecast System: GFS, ECMWF, NCEP-GFS, JMA, UK
 Regional Forecast Model: WRF-D, TWRF, NFS …
 Storm Scale NWP: LAPES-WRF, STAMAS-WRF …
 Ensemble Based NWP Forecast
 CWB WRF Ensemble Forecast System: 20 members
 TTFRI Ensemble Forecast System: 16-20 members
 Advanced Ensemble Transformed Forecast:
 ETQPF (Ensemble Typhoon QPF): QPF from a selected track
 And more…
 Subjective Adjustment by Senior Forecasters
 Recent Error Checking and Verification
 Forecaster’s Experiment
Concluding Remark
 Ensemble QPF Approach:
 Deterministic Ensemble QPF:
 Basic: Mean, median, variance, …
 Advanced: PM, newPM, ETQPF, …
 Probability of QPF (PQPF)
 Probability
 Most models need not probability but QPF
 Probabilistic but deterministic QPF (QPFP)
 Easy to make decision for Forecasters
 Probability inside
 Pattern-Recognition/Analog QPF on Observed Radar CV & WEPS
for (Extending) Nowcasting
 Better Pattern-Recognition method or strategy
 Fitting forecast among Radar QPF and Nowcasting
 Calibration or modification technology
 Storm Scale/ Radar/ Cloud resolved data assimulation and
prediction

Download Report