DA for watershed water quality forecasting

Ensemble Data Assimilation

for Watershed Water Quality

Forecasting

Sunghee Kim1,Hamideh Riazi1, D.-J. Seo1,

Changmin Shin2 ,Kyunghyun Kim2

1Dept. Of Civil Eng., The Univ. of Texas at Arlington, Arlington, TX, USA2Water Quality Control Center, National Institute of Environmental

Research, Incheon, Korea

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 1

DA for watershed water quality forecasting

In this presentation

• Water quality forecasting

– Needs, challenges for DA

• Maximum likelihood ensemble filter for

Hydrologic Simulation Program – Fortran

(MLEF-HSPF)

• Hindcasting experiment

• Operational implementation

• Conclusions and research questions

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 2

DA for watershed water quality forecasting

Why watershed water quality

forecasting?

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 3

DA for watershed water quality forecasting

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 4

base on Scenario #3

DA for watershed water quality forecasting

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 5

DA for watershed water quality forecasting

Operational WQ forecasting in

NIER, Korea

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 6

DA for watershed water quality forecasting

Operational WQ forecasting in

NIER, Korea

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 7

MLEF-

HSPF

DA for watershed water quality forecasting

DA for HSPF – Challenges

• High dimensionality

– A large number of control variables (28 state

variables, 31 model segments → 331 control

variables)

• Sparse observations

– In-river variables only

– Weekly only

• Nonlinear bio-physiochemical processes

• Nonlinear observations

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 8

DA for watershed water quality forecasting

MLEF-HSPF

• Based on MLEF (Zupanski, 2005)

– Model error added via state augmentation

• Formulated as a fixed-lag smoother

• Bias correction added in the observation equations to

remove/reduce systematic biases

– Conditional bias-penalized linear regression (Seo 2012)

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 9

Takkf

TTakkakk

Tkkakkf

kQkPMkP

kQkQkPMkPM

kQMkPMkP

)]()}([{)(

)()()}({)}({

)()()(

//,

/

///,

/,

,,

11

1111

11

21211

21

22121

211

11

DA for watershed water quality forecasting

Hindcasting experiment

• 2 catchments in the Nakdong River Basin, Korea, for

2008-2009

– Kumho, Banbyeon

• Assimilate weekly observations at the outlet and up to 4

interior monitoring stations

– Flow, TW, NH4, NO3, PO4, CHL-a, TN, TP, TOC, BOD,

DO

• Ensemble size of 9 based on sensitivity and eigenvalue

spectrum analysis (see poster by Riazi et al.)

• No. of control variables: 333 for Kumho, 316 for Banbyeon

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 10

DA for watershed water quality forecasting

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 11

(2,000 km2)

Nakdong River Basin (23,817 km2)

Meteorological stations

(2,000 km2)

DA for watershed water quality forecasting

Kumho, CHL-a, outlet only

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 12

RM

SE

(ug/l)

Analysis Day-1 Day-2 Day-3

Base

BC-Base

BC-DA

HSPF simulation without

Bias correction or DA

HSPF simulation with

bias correction

HSPF simulation with

bias correction and DA

15% 15% 18%

DA for watershed water quality forecasting

Banbyeon, DO, outlet only

Dec 12, 2013 AGU 2013 Fall Meeting 13

Analysis Day-1 Day-2 Day-3

RM

SE (

mg

/l)

Base

BC-Base

BC-DA

DA for watershed water quality forecasting

MSE skill score - Kumho

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 14-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

BOD CHL-a DO NO3 PO4 TW FLOW

Skill

Sco

re

Analysis DAY-1 DAY-2 DAY-3

DA for watershed water quality forecasting

Kumho

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 15

Reach

No.

Sampling

frequency

(sample

size)

Missing

observations

109 Weekly (96) Flow

118 Weekly (96) NO3, PO4

CHL-a in 2008

121 Weekly (96) NO3, PO4

124 Weekly (96) Flow

126 Weekly (96) Flow

127 Weekly (96) NO3, PO4

CHL-a in 2008

135 Weekly (96) ------

Reach 135: outlet

DA for watershed water quality forecasting

CHL-a at Reach 121

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 16

Obs from 1station (121)

assimilated

RM

SE

(ug/l)

Analysis Day-1 Day-2 Day-3 Analysis Day-1 Day-2 Day-3

Obs from 4 (121, 109, 118, 127)

stations assimilated

(NO3, PO4 observations missing)

RM

SE

(ug/l)

DA for watershed water quality forecasting

MLEF-HSPF

• Plugin module for the Water Quality Forecast System at

the National Institute of Environmental Research

(WQFS-NIER)

– FEWS-based

• Module components

– MLEF-HSPF program

– HSPF processor

• Interfaces MLEF-HSPF with HSPF

– MLEF-HSPF adapter

• Interfaces MLEF-HSPF with WQFS-NIER

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 17

DA for watershed water quality forecasting

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 18

Ensemble Data Assimilation:

MLEF

Hydrologic Simulation Program-Fortran

(HSPF)

Import in real time

MLEF-HSPF Adapters

HSPF Processor

• Weather

• Water quality

• Hydrologic data

WQFS-NIERExport forecasts for display

Set parameters

• Forecast frequency

• Length of lead time

• Model links

MLEF-

HSPF

module

DA for watershed water quality forecasting

Ongoing work

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 19

• Multi-catchment, multi-basin

evaluation

– Data analysis

– Hindcasting

• Operational implementation of MLEF-

HSPF

– Configuration for WQFS-NIER

River

basin

Number of

Catchment

Number of

monitoring stations

Han 9 86

Keum 7 21

Yeongsan 6 24

DA for watershed water quality forecasting

Conclusions & research questions

• DA is generally effective in improving accuracy in water

quality prediction

– Correction of model biases as part of the observation

equation is important

• MLEF handles nonlinear observation equations very well

• Improvement is larger for Banbyeon (natural) than Kumho

(urbanized)

• How underdetermined is the inverse problem? How to

reduce?

• Toward ensemble forecasting

– Verify and improve the quality (in particular, reliability and

spatio-temporal consistency) of updated ensemble ICs

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 20

DA for watershed water quality forecasting

Thank you

For more information:

sunghee@uta.edu

9/12/2014 CAHMDA-HEPEX/DAFOH Workshop, Austin, TX 21