an overview of the clever model
TRANSCRIPT
An Overview of the CLEVER Model:the real-time flood forecasting system for BC–– Challenges, science, operations, history and uncertainties
Charles LuoBC River Forecast Centre, FLNRORD
April 29, 2021
1
2
Contents1. Why the CLEVER Model – Challenges of real-time flood forecasting in BC 2. Model structure3. Science: Distributed open channel routing – An improved kinematic wave
scheme4. Science: Lumped watershed routing – An improved temperature-index
snowmelt model5. Codes and user interfaces6. Input data and data assimilation7. Model calibration – flexibility to adapt to climate changes8. Producing forecasts – informative and easy to read9. Model history – developing, expanding and upgrading10. Post freshet review of model performance - statistics and flooding events11. Modeling uncertainties – understanding limitations12. Types of modeled stations
13
Area (km2):----------------------------------------British Columbia: 947,900=91% of the following states:-------------------------------------Washington: 184,827 Oregon: 254,806 Idaho: 216,443 Montana: 380,800 -------------------------------------Total: 1,036,876
Model Area: 700,401 (km2, 74% of BC)
Model stations: 311
1.Why the CLEVER Model – Challenges of real-time flood forecasting in BC
Challenge 1.Supper large areas
Total length of rivers in BC = 42,150 km > earth’s circumference
14
0200400600800
100012001400160018002000
02-12 02-13 02-14 02-15 02-16 02-17
PEACE RIVER AT HUDSON HOPE (07EF001)
Disc
harg
e(m
3 /s)
Date (MM-DD in 2016)
Q increasing by 1200 m3/s in 3 hours!!!
Highly regulated basins
Lakes
1.Why the CLEVER Model – Challenges of real-time flood forecasting in BC
Water level only stations
Climate pattern variability Human/natural disturbances
Mountain pine beetle infests
Wildfires
Sparce climate/flow stations
Challenge 2. Tremendous heterogeneity
16
1.Why the CLEVER Model – Challenges of real-time flood forecasting in BC
Stantec Consulting Ltd, 2014. River Forecast Performance Measures Development ProjectFinal Report, Alberta Environment and Sustainable Resource Development, Edmonton, AB
<=4
Challenge 3. Very limited resources
European CERN Computer
17
1.Why the CLEVER Model – Challenges of real-time flood forecasting in BC
Observed
Climate & Flow
Data
Latest
Challenge 4. Very limited time for modeling (about 90 minutes each day)
18
1.Why the CLEVER Model – Challenges of real-time flood forecasting in BC
Challenge 1.Supper large areas
Challenge 2. Tremendous heterogeneity
Challenge 3. Very limited resources
Challenge 4. Very limited time for modeling (about 90 minutes each day)
Channel Links Evolution EfficientRouting (CLEVER) Model
19
2. Model structure – watershed simplification
4
32
1
Flow gauge station
Watershed
Watershed outlet
River
River
River
4
32
1
Sub-basin 1
Sub-basin 2
Sub-basin 3
Sub-basin 4
Channel link
Channel link
4
32
1
20
2. Model structure – Watershed routing
21
2. Model structure – Open channel routing
Space (x)
Tim
e (t) x=i-1 x=i
t=j
t=j-1
S0
S0
Ai-1, j
Qi-1, j
Ai-1, j-1
Qi-1, j-1
Ai, j
Qi, j
Ai, j-1
Qi, j-1
Improved kinematic wave scheme
22
2. Model structure – Model flow chart
Read watershed parameters
Read observed
and forecast climate data
Read observed flow data
Groundwater release, evapo-
transpirtation and infiltration
Daily climate data
distribution to hourly data
An improved temperature-
index snowmelt model
Net water input
Instantaneous unit hydrograph
Open channel routing from basin
center to outlet
Lake basin routing:Level – Volume –
Discharge
Lumped Watershed Routing Sub-model
Regulated basin routing:
Extending observed trend
Distributed Open Channel Routing Sub-model
Err<=Min
No
Calculate flux limiter φ
Calculate V
Calculate A
Yes
An improved
kinematic wave
scheme
Iteration
Calculate boundary conditions
Calculate initial
conditions
Read inflowRead
channel link parameters
Call
24
3. Science: Distributed open channel routing –– An improved kinematic wave scheme
0200400600800
100012001400160018002000
02-12 02-13 02-14 02-15 02-16 02-17
PEACE RIVER AT HUDSON HOPE (07EF001)
Disc
harg
e(m
3 /s)
Date (MM-DD in 2016)
Highly regulated hydrograph
𝜕𝜕𝑄𝑄𝜕𝜕𝑥𝑥
+𝜕𝜕𝐴𝐴𝜕𝜕𝑡𝑡
= 0
𝑆𝑆0 =𝑛𝑛2𝑄𝑄2
𝐴𝐴2𝑅𝑅 ⁄4 3
𝐴𝐴𝑖𝑖,𝑗𝑗 =∆𝑡𝑡 𝑄𝑄𝑖𝑖−1,𝑗𝑗 + 𝑄𝑄𝑖𝑖−1,𝑗𝑗−1 − 𝑄𝑄𝑖𝑖,𝑗𝑗−1 + ∆𝑥𝑥 𝐴𝐴𝑖𝑖,𝑗𝑗−1 + 𝐴𝐴𝑖𝑖−1,𝑗𝑗−1 − 𝐴𝐴𝑖𝑖−1,𝑗𝑗
∆𝑡𝑡𝑉𝑉𝑖𝑖,𝑗𝑗 + ∆𝑥𝑥
𝑉𝑉𝑖𝑖,𝑗𝑗 =1𝑛𝑛
𝑆𝑆0𝑅𝑅𝑖𝑖,𝑗𝑗⁄2 3𝑄𝑄𝑖𝑖.𝑗𝑗 = 𝑉𝑉𝑖𝑖,𝑗𝑗𝐴𝐴𝑖𝑖,𝑗𝑗
𝐴𝐴𝑖𝑖,𝑗𝑗𝑘𝑘 =
∆𝑡𝑡 𝑄𝑄𝑖𝑖−1,𝑗𝑗 + 𝑄𝑄𝑖𝑖−1,𝑗𝑗−1 − 𝑄𝑄𝑖𝑖,𝑗𝑗−1 + ∆𝑥𝑥 𝐴𝐴𝑖𝑖,𝑗𝑗−1 + 𝐴𝐴𝑖𝑖−1,𝑗𝑗−1 − 𝐴𝐴𝑖𝑖−1,𝑗𝑗
∆𝑡𝑡 𝑉𝑉𝑖𝑖,𝑗𝑗𝑘𝑘−1 + ∆𝑥𝑥
Finite difference with the Preissmann scheme:
500
1000
1500
2000
2500
0 20 40 60 80 100 120
OBS (07FD010) LUO_SCN1
Disc
harg
e(m
3 /s)
Time (hour)
𝑟𝑟𝑖𝑖,𝑗𝑗 =𝑄𝑄𝑖𝑖−1,𝑗𝑗−1 − 𝑄𝑄𝑖𝑖−1,𝑗𝑗−2
𝑄𝑄𝑖𝑖−1,𝑗𝑗 − 𝑄𝑄𝑖𝑖−1,𝑗𝑗−1𝜑𝜑 𝑟𝑟 = 𝑚𝑚𝑚𝑚𝑥𝑥 0,𝑚𝑚𝑚𝑚𝑛𝑛 1, 𝑟𝑟
𝐴𝐴𝑖𝑖,𝑗𝑗𝑘𝑘 =
∆𝑡𝑡𝑄𝑄𝑖𝑖−1,𝑗𝑗 + ∆𝑥𝑥𝐴𝐴𝑖𝑖,𝑗𝑗−1
∆𝑡𝑡 𝑉𝑉𝑖𝑖,𝑗𝑗𝑘𝑘−1 + ∆𝑥𝑥
+ 𝜑𝜑 𝑟𝑟𝑖𝑖𝑗𝑗∆𝑡𝑡 𝑄𝑄𝑖𝑖−1,𝑗𝑗−1 − 𝑄𝑄𝑖𝑖,𝑗𝑗−1 + ∆𝑥𝑥 𝐴𝐴𝑖𝑖−1,𝑗𝑗−1 − 𝐴𝐴𝑖𝑖−1,𝑗𝑗
∆𝑡𝑡 𝑉𝑉𝑖𝑖,𝑗𝑗𝑘𝑘−1 + ∆𝑥𝑥
Minmod Flux limiter
Kinematic wave of Saint Venant Equation:
V unknow -> Iteration:
25
3. Science: Distributed open channel routing –– An improved kinematic wave scheme
Luo, C. 2021. Comparing Five Kinematic Wave Schemes for Open-Channel Routing for Wide-Tooth-Comb-Wave Hydrographs. Journal of Hydrologic Engineering, ASCE, 26 (4).
28
4. Science: Lumped watershed routing – An improved temperature-index snowmelt model
Time Efficiency
Simplex=y
Temperature-index method: 𝑴𝑴 = 𝑴𝑴𝒇𝒇 𝑻𝑻𝒊𝒊 − 𝑻𝑻𝒃𝒃M: snowmelt𝑀𝑀𝑓𝑓: melt factor𝑇𝑇𝑖𝑖: air temperature 𝑇𝑇𝑏𝑏: base temperature snow starts to melt
A forest stand
A small basin such as:Kanaka Creek
A=49 km2Water balance equation:𝑊𝑊 = 𝑅𝑅 + 𝑀𝑀 + 𝐺𝐺 − 𝐸𝐸 − 𝐼𝐼
29
4. Science: Lumped watershed routing – An improved temperature-index snowmelt model
FRASER RIVER AT MCBRIDEA = 5,164 km2
Temperature-index on watershed-scale using hourly time step:
𝑴𝑴 = 𝒄𝒄𝒂𝒂𝒄𝒄𝒅𝒅𝑴𝑴𝒇𝒇 𝑻𝑻𝒊𝒊 − 𝑻𝑻𝒃𝒃 𝜷𝜷
𝑐𝑐𝑎𝑎: snow covering correction factor during the snowpack receding period (<=1):
𝑐𝑐𝑎𝑎 =𝑆𝑆𝑊𝑊𝐸𝐸𝑖𝑖
𝑆𝑆𝑊𝑊𝐸𝐸𝑚𝑚𝑎𝑎𝑚𝑚
𝛼𝛼
𝑐𝑐𝑑𝑑: Ordinal day correction factor (<=1)𝛼𝛼: snowpack covering area receding power𝛽𝛽: temperature power, which could be >,=,<1
0
0.2
0.4
0.6
0.8
1
1.2Cd_BAKR Cd_SJBC Cd_WSRD Cd_NAZK
Ord
inal
day
cor
rect
ion
fact
orCd
Date (YYYY-MM-DD)
Energy for melt is nonlinear to T
T usually lags and attenuate in daytime
Hourly time step shorter than time needs for snow to melt as estimated
Large portion of radiation absorbed by vegetation for photosynthesis
Whentemperatures increase
significantlyand rapidly
Underestimateslope of rise
(flatter)
0
2
4
6
8
10
12
14
0 10 20 30 40
β=0.75 β=1 β=1.1
Snow
mel
t (m
m)
T (oC)
Mf=0.25, ca=1, cd=1
Overestimateslope of rise
(steeper)
Improved
32
5. Codes and user interface
5 modules12,492 lines
33
5. Codes and user interfaces
Model control interface
34
5. Codes and user interfaces
Watershed interface
Input data recording
area
Watershed parameter panel Hydrograph area
Output data recording
area
35
6. Input data and data assimilation
ECCC’s MPOML climate dataObserved
climate data
ECCC climate stations
BC fire weather stations
Automated snow weather
stations
Daily max & min temperatures & 24 hour total rainfall
36
6. Input data and data assimilation
Forecast climate data
Canadian Meteorological Centre’s (CMC) Numerical Weather Prediction (NWP) -GRIB2 format
◾Regional Deterministic
Prediction System -RDPS (dx=10 km, dt=3h, Time=0 to
62h)
Locations of 368 climate stations
(about 20 minutes)
Downscaleto
◾Global Deterministic
Prediction System -GDPS (dx= 25 km, dt =3h, Time used= 63
to 240h)
Map of CMC’s10 forecasts of daily average T and daily Phttp://bcrfc.env.gov.bc.ca/freshet/FORECAST_CMC_MAP.pdf
161 files to download (about 20 minutes)
37
6. Input data and data assimilation
Observed Hydrometric
data
Water Survey of Canadian DataMart hydrometric data
Hourly+
Daily
Different time step
Obviously wrong data
Irregularly missing
Completely missing
38
6. Input data and data assimilation
In the model:Dily climate data are distributed to
hourly data
40
7. Model calibration – flexibility to adapt to climate changes
42
7. Model calibration – flexibility to adapt to climate changes
Hydrographs
Observed & estimated discharges/water levels
Model parameters
43
8. Producing forecasts – informative and easy to read
46
8. Producing forecasts – informative and easy to read
330m3/s
Forecast hydrograph
56
9. Model history – developing, expanding and upgrading
The Channel Links Evolution Efficient Routing (CLEVER) Model
2013 Planning, researching and coding the first version of the CLEVER Model(M)
32 stn
2014 Test running the CLEVER Model (Moderate water year)
51 stn
2015 Starting operational running the CLEVER Model (Low water year)
71 stn
2016 Operational running the CLEVER Model (Very low water year)
74 stn
2017 Significant increase of modeled stations (Moderate water year)
102 stn
2018 Operational running the CLEVER Model (High water year)
108 stn
2019 Providing CSV files of hourly forecast for users to download (Low water y)
119 stn
2020 Significant expanding modeling scope (Extensive flooding year)
266 stn
2021
1. Massive upgrading, including feature allowing multi-users to run the model in parallel to improve time efficiency.
2. First paper about the CLEVER Model published in an international peer reviewed journal (J. Hydro. Eng., ASCE)
311 stn
10. Post freshet review of model performance– statistics and flooding events
𝐶𝐶𝑒𝑒 = 1 −∑𝑗𝑗=1𝑚𝑚 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜
𝑗𝑗 − 𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚𝑗𝑗 2
∑𝑗𝑗=1𝑚𝑚 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜𝑗𝑗 − 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜
2
𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜 =1𝑚𝑚�𝑗𝑗=1
𝑚𝑚
𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜𝑗𝑗
Coefficient of model efficiency (𝐶𝐶𝑒𝑒)
Coefficient of determination (𝐶𝐶𝑑𝑑)
Percentage volume difference (𝑑𝑑𝑉𝑉)
𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚 =1𝑚𝑚�𝑗𝑗=1
𝑚𝑚
𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚𝑗𝑗
𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜2 =1𝑚𝑚�𝑗𝑗=1
𝑚𝑚
𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜𝑗𝑗 2
𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚2 =1𝑚𝑚�𝑗𝑗=1
𝑚𝑚
𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚𝑗𝑗 2
𝐶𝐶𝑑𝑑 = 1 −∑𝑗𝑗=1𝑚𝑚 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜
𝑗𝑗 − 𝑚𝑚 � 𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚𝑗𝑗 + 𝑏𝑏
2
∑𝑗𝑗=1𝑚𝑚 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜𝑗𝑗 − 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜
2
𝑚𝑚 = ��𝑃𝑃 − 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜 � 𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚 𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚2 − 𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚2
𝑏𝑏 = 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜 − 𝑚𝑚 � 𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚
𝑑𝑑𝑉𝑉 = 100 × ⁄𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚 − 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜
�𝑃𝑃 =1𝑚𝑚�𝑗𝑗=1
𝑚𝑚
𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜𝑗𝑗 � 𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚
𝑗𝑗
Statistic analysis for model calibration
Statistic analysis for forecasts
Relative mean absolute error (𝐸𝐸𝑟𝑟𝑎𝑎)
𝐸𝐸𝑟𝑟𝑎𝑎 = �100 ×1𝑚𝑚�𝑗𝑗=1
𝑚𝑚
𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚𝑗𝑗 − 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜
𝑗𝑗 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜
𝑟𝑟2 =∑𝑗𝑗=1𝑚𝑚 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜
𝑗𝑗 − 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜 𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚𝑗𝑗 − 𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚
2
∑𝑗𝑗=1𝑚𝑚 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜𝑗𝑗 − 𝑄𝑄𝑜𝑜𝑏𝑏𝑜𝑜
2∑𝑗𝑗=1𝑚𝑚 𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚
𝑗𝑗 − 𝑄𝑄𝑜𝑜𝑖𝑖𝑚𝑚2
r squared (𝑟𝑟2)
62
Statistic analysis:
63
10. Post freshet review of model performance– statistics and flooding events
After freshet review (AFR) reports as of 2017:
1. Changes of modeling scope2. Statistics of model calibration and forecasts3. Model performance during flooding events4. Recommendations for model improvements
65
10. Post freshet review of model performance– statistics and flooding events
Forecast errors of FRASER RIVER AT HOPE (08MF005)
0
5
10
15
20
Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 Day 8 Day 9 Day 10
Forecast Error (%)
2015 2016 2017 2018 2019 2020 Ave
Forecast Error (%) StatisticsYear Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 Day 8 Day 9 Day 10 AveErr RSQU2015 0.6 2.1 3.5 4.2 5.1 6.2 7.8 9.0 10.4 11.2 6.0 0.7382016 1.0 2.6 3.9 5.1 6.4 7.7 9.8 12.1 14.2 15.1 7.8 0.4442017 0.9 2.6 3.8 4.7 5.2 5.6 6.1 6.5 7.6 9.1 5.2 0.7722018 1.4 4.0 6.2 7.9 9.2 10.5 12.1 13.3 14.7 16.0 9.5 0.6062019 1.1 3.2 5.0 6.3 6.9 8.1 9.1 10.0 12.2 13.1 7.5 0.5542020 1.2 3.8 5.6 7.3 8.9 10.6 12.2 13.9 16.1 17.6 9.7 0.497Ave 1.0 3.0 4.7 5.9 7.0 8.1 9.5 10.8 12.5 13.7 7.6 0.602
67
10. Post freshet review of model performance– statistics and flooding events
Review of Granby May 10, 2018 Flood (3:55 p.m. PST) (Observed Q peak = 524 m3/s)
68
10. Post freshet review of model performance– statistics and flooding events
http://bcrfc.env.gov.bc.ca/freshet/cleverm_ref/ChilcotinFlood2019July_lr.pdf
(78 pages)
Download April 24, 2021
72
11. Modeling uncertainties – understanding limitations
First category of modeling uncertaintiescoming input data
Observed flow data: gauge issues and missing data
Observed climate data: missing and
interpolation/ extrapolation
Forecast climate data (ECCC CMC NWP Grib2
Regional MD 10km/ global MD 25km) and
downscaling to clim. st
Missing data
75
11. Modeling uncertainties – understanding limitations
Second category of modeling uncertaintiescoming from the model itself
Model intrinsic limitations
Simplifying a basin into a single node
Using daily data to distribute into hourly
Lack of overland flow routing
Simplifying hydrologic cycle: Evp, Inf, GD, SM
Model operational limitations
Parameters calibrated in previous years may have changed without
recent hydrologic indicators
Watershed routing
Open channel routing
Kinematic wave simplification of SV Eq.
Finite difference solution of KW Eq.
Simplification of cross-section into a rectang.
Lack of data of channel dimensions (42,150 km)
Too many parameters to calibrate – wrong combination of
parameters for correct results
Too short time for model calibration – lower accuracy
Modelers’ experience
76
12. Types of modeled stations
The only type of stations included in the model when developed and in early stage of operation (watersheds were split originally based on WSC stations)
1. After April 2018 West Road River/Nazko River flood, the inactive WSC station NAZKO RIVER ABOVE MICHELLE CREEK (08KF001) was added because of flooding concerns.
2. After July 2019 Chilcotin River Flood, WSC eliminated the CHILCOTIN RIVER BELOW BIG CREEK (08MB005), which was kept in the model as an inactive station for better model calibration.
3. PEACE RIVER AT HUDSON HOPE (07EF001) became inactive in 2019 and removed from the model in 2020. It was restored in the model as an inactive station in 2021 for better model calibration.
4. Because of flooding concerns, the SOMASS RIVER NEAR ALBERNI (INACTIVE STATION) (08HB017) was included when VI basins were included in the model as of 2020.
The only Non-WSC station is the LILLOOET RIVER FSR (08MG00A) (BC station ID: 08MG0001), an important station for Pemberton.
77
Liard River,
A=104,000 km2
Split into 20 sub-basins with routing stations and YT stations
12. Types of modeled stations
Estimated station