Evaluating hydrological model structure using tracer data within a multi-model framework

1
Evaluating hydrological model structure using
tracer data within a multi-model framework
H31F-1227
Hilary McMillan1, Doerthe Tetzlaff2, Martyn
Clark3, Chris Soulsby2
1 National Institute of Water and Atmospheric
Research, New Zealand.
2 School of Geosciences, University of Aberdeen
3 National Centre for Atmospheric Research,
Boulder, Colorado
Aims
Water-Tracking Method
Models with physically realistic structures are
needed to produce good forecasts under a wide
range of conditions
But very different model structures can produce
similar hydrographs

1. Use tracer data to choose between model
   structures with similar dynamics
2. Show how tracer response is affected by
   interaction of model structure, parameters and
   mixing assumptions

FUSE multi-model framework was modified to track
distributions of water age in each store and
flux. Outflow age distributions, transit time
distributions and tracer dynamics can be derived.
-- The story --
Flow (mm)
FUSE (Framework for Understanding Structural
Errors) allows modular testing of popular
hydrological model components
Time
To choose between model structures we use
diagnostic tests which target individual model
components using data types such as flow, soil
moisture, and here tracers
Case Study Loch Ard, Scotland
Results
Tracer Simulation
Seasonality
Sensitivity
Loch Ard Catchment
Structure vs Calibration Some parameters (e.g.
upper soil zone depth) control transit times
equally with model structure
Model virtual experiments allow transit time
behaviour to be explored We tested which transit
time characteristics lead to good model
performance
At Loch Ard a FUSE model could be designed which
simulated both runoff and tracer dynamics Key
choices were a structure with single upper zone
variable and Topmodel lower zone architecture
Transit Time Distributions
Vary Lower Zone Size
Vary Upper Zone Size
Loch Ard Burn 10 is a small (0.9km2) catchment
forested with Sitka spruce. Soils are poorly
drained gleys and storm runoff is dominated by
the upper soil horizons. Tree roots and exposed
bedrock allow deeper recharge.
Model is more sensitive to store depth when store
response is more nonlinear
Time-varying Mean Transit Times
Frequency
Linear Tank
2 Linear Tanks
Topmodel
Models which perform well have strong seasonal
variation in MTT
Time (days)
Model simulates flow and tracers
Modelled tracer series

Time (days)
Time (days)
Models with single upper zone variable
i
i
12 years of data was used rainfall, flow and
weekly samples of chloride. Chloride in rainfall
originates from sea-salt and the concentration
varies seasonally due to wind speed and direction
Nash Score
Measured Flow
Rain (mm)
Flow (mm)
Chloride (mg/L)
Lower Zone Size (mm)
Upper Zone Size (mm)
Model simulates flow but not tracers
Time
Models with split upper zone variables
High MTT in summer
Low MTT in winter
Mixing Assumptions Does saturation excess flow
mix with soil water?
Flow (mm)
D
i
Date
Seasonal Transit Time Distributions
TTD with variable mixing
Chloride (mg/L)
More mixing
Flow partitioning between surface and soil water
was found to have only a small effect on transit
times so the simplifying assumption of no mixing
was acceptable
Transit Time Simulation
Frequency
Steady state transit time distributions
Less mixing
Differences in tracer response could be explained
by differences in model transit time distribution
Time (days)
of flow
Transit time distributions for fast flow pathways
(lt30 days) depend strongly on catchment wetness.
At these timescales we shouldnt assume the TTD
is stationary. At timescales gt30 days,
seasonality is less important.
Less mixing
More mixing
Contact h.mcmillan_at_niwa.co.nz