Вы находитесь на странице: 1из 20

HYDROLOGICAL PROCESSES

Hydrol. Process. 26, 3012–3031 (2012)


Published online 4 January 2012 in Wiley Online Library
(wileyonlinelibrary.com) DOI: 10.1002/hyp.8247

Sensitivity of the Weather Research and Forecasting (WRF)


model to downscaling ratios and storm types in rainfall
simulation
Jia Liu,1,2* Michaela Bray3 and Dawei Han2
1
State Key Laboratory of Simulation and Regulation of Water Cycle in River Basin, China Institute of Water Resources and Hydropower Research,
Beijing 100038, China
2
Water and Environmental Management Research Centre, Department of Civil Engineering, University of Bristol, Bristol BS8 1TR, UK
3
Hydro-environmental Research Centre, School of Engineering, Cardiff University, Cardiff CF24 0DE, UK

Abstract:
Accurate information of rainfall is needed for sustainable water management and more reliable flood forecasting. The advances
in mesoscale numerical weather modelling and modern computing technologies make it possible to provide rainfall simulations
and forecasts at increasingly higher resolutions in space and time. However, being one of the most difficult variables to be
modelled, the quality of the rainfall products from the numerical weather model remains unsatisfactory for hydrological
applications. In this study, the sensitivity of the Weather Research and Forecasting (WRF) model is investigated using different
domain settings and various storm types to improve the model performance of rainfall simulation. Eight 24-h storm events are
selected from the Brue catchment, southwest England, with different spatial and temporal distributions of the rainfall intensity.
Five domain configuration scenarios designed with gradually changing downscaling ratios are used to run the WRF model with
the ECMWF 40-year reanalysis data for the periods of the eight events. A two-dimensional verification scheme is proposed to
evaluate the amounts and distributions of simulated rainfall in both spatial and temporal dimensions. The verification scheme
consists of both categorical and continuous indices for a first-level assessment and a more quantitative evaluation of the
simulated rainfall. The results reveal a general improvement of the model performance as we downscale from the outermost to
the innermost domain. Moderate downscaling ratios of 1:7, 1:5 and 1:3 are found to perform better with the WRF model in
giving more reasonable results than smaller ratios. For the sensitivity study on different storm types, the model shows the best
performance in reproducing the storm events with spatial and temporal evenness of the observed rainfall, whereas the type of
events with highly concentrated rainfall in space and time are found to be the trickiest case for WRF to handle. Finally, the
efficiencies of several variability indices are verified in categorising the storm events on the basis of the two-dimensional rainfall
evenness, which could provide a more quantitative way for the event classification that facilitates further studies. It is important
that similar studies with various storm events are carried out in other catchments with different geographic and climatic
conditions, so that more general error patterns can be found and further improvements can be made to the rainfall products from
mesoscale numerical weather models. Copyright © 2011 John Wiley & Sons, Ltd.

KEY WORDS rainfall simulation; numerical weather modelling; Weather Research and Forecasting (WRF) model; sensitivity
study; downscaling ratio; storm type
Received 18 November 2010; Accepted 28 July 2011

INTRODUCTION dense enough or inappropriately located to provide an


accurate and representative measurement of the catchment
As a driving force for runoff, rainfall plays an important
rainfall. The problem is even more serious with ungauged
role in the hydrological cycle as well as in climatic and
catchments. The global numerical weather prediction
ecological systems. In real-time flood forecasting, the
(NWP) systems could provide global forecast and
forecasted rainfall is needed for an extension of the flood
assimilated climate products, which are potentially useful
forecast lead time, which is of particular importance in
sources for the rainfall data. However, the spatial and
small- and medium-sized basins with quick catchment
temporal resolutions of the global products are too coarse
responses (Brath et al., 1988; Cluckie and Han, 2000).
for hydrological use for the purpose of, for example, flood
Meanwhile, for a sustainable management of the local
forecasting and water management. In that case, the so-
water resources, it is important to have reliable and
called downscaling procedure is needed to bridge the scale
accurate information of the catchment rainfall at as fine a
gap between the large-scale global NWP domains and the
resolution as possible. In addition, for most catchments
catchment-size domains for hydrological use, which is
all over the world, the position of rain gauges is either not
realised by the mesoscale NWP models.
Nowadays, the advances in mesoscale numerical wea-
ther modelling and the modern computing technologies
*Correspondence to: Jia Liu, 0.36 Queen’s Building, University Walk,
Clifton, Bristol BS8 1TR, UK. make it possible to provide rainfall simulations and
E-mail: Jia.Liu@bristol.ac.uk forecasts at increasingly higher resolutions by downscaling

Copyright © 2011 John Wiley & Sons, Ltd.


SENSITIVITY OF WRF MODEL IN RAINFALL SIMULATION 3013

products from the global NWP systems (Xuan et al., 2005). over Mumbai, India, with the amount of the 24-h rainfall
The Weather Research and Forecasting (WRF) model is the exceeding 944 mm, the results of which showed a high
latest generation of the mesoscale NWP models that serves sensitivity to the grid spacing with finer grids leading to higher
both operational and research communities. It has been rainfall.
developed under the joint efforts of the scientific All the previous studies generally indicate an improve-
community on the basis of the experience of its ment of the rainfall simulation or prediction with the
predecessors, such as the widely used MM5 (the Fifth increase of the spatial resolution. However, it might be
Generation Penn State/NCAR Mesoscale Model). The asked to what extent could the WRF model provide further
performance of the WRF model has been validated through improved results? Or in other words, should we down-
increasingly more applications in the meteorological area scale the global data to as fine a resolution as possible
(e.g., Done et al., 2004; Davis et al., 2006; Kain et al., even at the expense of the largely increased modelling
2006; Etherton and Santos, 2008; Mölders and Kramm, time? Downscaling from the coarse resolution global data
2010). Since rainfall process is triggered by large-scale air to a finer resolution at the catchment scale is normally
motions and local convections that are difficult to model, realised by running a mesoscale NWP model with multiple
rainfall simulation and forecasting using mesoscale NWP nested domains of decreased domain size and grid
models has long been recognised as one of the most spacing. This poses another interesting question: Does
difficult tasks among all the meteorological variables (Toth the downscaling ratio have some underlying impact on the
et al., 2000). Recent studies have shown that the WRF performance of the WRF model? Essentially, the down-
model has good potential for capturing some features of scaling issue is no more than setting the initial condition
rainfall, for example, the rainfall timing, location and (IC) and the lateral boundary condition (LBC) for the
evolution; however, for capturing the amount of rainfall, mesoscale NWP model. With the domain size and the
the results are still not ideal (Chang et al., 2009; Hong and location of the nested system fixed, what downscaling ratio
Lee, 2009; Shem and Shepherd, 2009; Chen et al., 2010). will ensure the ideal results in the objective (normally the
Besides the uncertainties in the initial conditions, the innermost) domain? How to improve the efficiency and
limited knowledge regarding the rainfall process and the effectiveness of the downscaling process by adjusting the
problematic cloud microphysics (Fowle and Roebber, downscaling ratio? For MM5, the fixed ratio of 1:3 is
2003; Fritsch and Carbone, 2004), reasons for the necessary for any two-way nesting configuration (i.e.
unsatisfactory performance of the WRF model in rainfall feedback from the inner domain to the outer domain),
modelling have also been found to be related to the whereas in the WRF model, the domain setting is more
sensitivity of the model performance concerning the flexible and there is no longer any restriction on the
configurations of the domain size and the grid spacing downscaling ratio. However, up until now, most studies
(Knievel et al., 2004). As early as its previous generation, still use the recommended ratio of 1:3 in the WRF model.
the MM5 model, attentions have been drawn to the Will this always be the most appropriate ratio for the
relationship between the model accuracy and the mesoscale NWP models? By intuition, we assume that the
resolution of the domain. A flood event caused by a serious downscaling process should be carried out gradually from
Pacific storm in southwest Washington and northwest the outermost to the innermost domain; thus, a moderate
Oregon was simulated in the study of Colle and Mass downscaling ratio is probably the best choice to obtain
(2000) at 36-, 12-, 4- and 1.33-km horizontal resolutions more appropriate downscaled results. Hence, could alter-
using the MM5 model. It was found that there was a natives or more appropriate downscaling ratios exist other
significant improvement in the rainfall forecasts as the grid than the most commonly used ratio of 1:3? Such questions
spacing was decreased from 36 to 4 km, but the increasing regarding the domain configurations for the WRF model
resolution from 4 to 1.33 km did not produce a significant need to be answered before we can perform efficient
improvement across the entire domain, with only some downscaling of the forecasted or assimilated products from
improvements in a few specific regions. Knievel et al. the global systems.
(2004) evaluated the rainfall simulation using a prototype Because the quantity of rainfall is one of the most
version of the WRF model for conterminous United States difficult variables for a numerical weather model to handle,
during the summer period. The results showed that most studies only focused on the synoptic analysis of some
simulations at different resolutions did produce different specific storm events with large intensities. Therefore, there
results with the 4-km results being superior to the 22- and is a lack of quantitative verification of the model
10-km results. Wang et al. (2006) investigated the nesting performance on rainfall modelling, and no sensitivity
techniques of the WRF model on the mesoscale simulation analysis is carried out regarding different storm types. In
of a Meiyu front severe rainfall storm, which happened in the this study, eight storm events with the same durations of
Huaihe river basin of China. The results indicated that the 24 h are selected from the Brue catchment, southwest
simulated rainfall on a finer resolution domain was closer to England. These are used to perform a sensitivity study of
the real rainfall and that there was an obvious improvement the performance of the WRF model to different domain
in the simulated distribution of the storm area and the settings and various storm types. A two-dimensional
position of the storm centre. Change et al. (2009) coupled verification scheme is proposed to evaluate the perform-
the WRF model with different land surface models to ance of the WRF model in both spatial and temporal
simulate a record-breaking heavy rain event that occurred dimensions. The verification scheme includes a first-level

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
3014 J. LIU, M. BRAY AND D. HAN

verification by using the categorical indices and a second- The model is viable across scales, ranging from metres (for
level verification with the continuous indices to further ideal cases) to thousands of kilometres. The sigma coordinates
evaluate the model performance more quantitatively. The are adopted to describe the vertical pressure levels, and two-
sensitivity analysis is divided into two parts. First, five way nesting is allowed for the interaction between the
domain configuration scenarios are designed with gradually mother and the child domains. In this study, all domains are
changing downscaling ratios and different grid spacing, composed of 28 vertical pressure levels with the top level set
which are then applied to the eight storm events to find out at 50 hPa. The model initial and lateral boundary conditions
the best scenario with the most appropriate downscaling are derived from the ECMWF 40-year reanalysis (ERA-40)
ratios. The performance of the WRF model from different data with the improved resolution of 1  1 and updated
domains with a certain scenario is also compared so as to every 6 h.
explore the improvement of the downscaling results with The WRF model has options for different physical
respect to the increase of the grid spacing or the decrease of parameterisations such as the boundary layer, the convec-
the spatial resolution. In the second part of the sensitivity tion and radiation schemes, and so forth. To focus on the
analysis, the eight events are categorised into four main effect of the domain configuration and the downscaling
storm types on the basis of the evenness of the spatial and ratio, the most extensively used parameterisations are
temporal distributions of the rainfall observations. The applied in this study for all the numerical experiments. The
performance of the WRF model is investigated and further main physics packages are the WRF single-moment three-
compared on the basis of different storm types. class microphysics scheme (Hong et al., 2004), the new
The purpose of this study is to search for some general Kain–Fritsch cumulus parameterisation scheme (Kain,
guidance on the domain settings (including the downscal- 2004), the Yonsei University planetary scheme for the
ing ratio, the grid spacing and the domain size) of the WRF planetary boundary layer (Hong et al., 2006), the Dudhia
model and to provide an insight into the limitations of the shortwave radiation scheme (Dudhia, 1989) and the rapid
model functions in reproducing different types of storm radiative transfer model long-wave radiation scheme
events. It should be mentioned that although the mesoscale (Mlawer et al., 1997). Other physics options include the
NWP models like WRF have been widely investigated by Monin–Obukhov scheme (Monin and Obukhov, 1954) for
meteorologists, for example, for the boundary condition the description of the surface layer and the Pleim-Xiu Land
effect or the choice of the model physics, and so forth, this Surface Model (Xiu and Pleim, 2001) to present the land
article addresses the sensitivity of the WRF model from the surface physics. The advection orders are chosen as the
viewpoint of the hydrological community who might fifth for the horizontal advection and the third in the
simply be the user of the rainfall products, in the aspects vertical direction. It should be noted that in this study the
of raising their awareness of the uncertainties in the NWP cumulus parameterisation is not used in the innermost
rainfall products, choosing appropriate domain settings and domain where the convective rainfall generation is
mitigating the inefficiency of WRF rainfall downscaling for assumed to be explicitly resolved.
hydrological applications. The 40-year reanalysis data from
the European Centre for Medium-range Weather Forecasts
(ECMWF) are used to drive the WRF model for the 24-h Model verification methods
rainfall simulations of the eight events. The reanalysis data To fully evaluate the performance of the WRF model, a
have been conditioned by a wide range of remote-sensed two-dimensional verification scheme is proposed. Two
observations (e.g., data from weather balloon, radio sonder, types of traditional verification indices, that is, the
satellite and ground based measurements, etc.), which have categorical and the continuous indices (Stanski et al.,
become one of the most important data sources for 1989; Jolliffe and Stephenson, 2003; Wilks, 2006), are
scientific and application communities (Betts et al., 2005; calculated in both spatial and temporal dimensions. The
Graversen et al., 2007). Because of the similar configu- categorical verification indices include the probability of
ration steps of the WRF model, the main conclusions from detection (POD), the frequency bias index (FBI), the false
this article are also transferable for the usage of the forecast alarm ratio (FAR) and the critical success index (CSI). The
products from the global NWP systems. continuous indices are chosen as the root mean square error
(RMSE), the mean bias error (MBE) and the standard
deviation (SD). The calculation of the categorical indices is
WRF MODEL CONFIGURATION AND based on the rainfall contingency table, as shown in
VERIFICATION Table I. Then the categorical indices can be obtained
Model description and configuration
The numerical experiments in this study are conducted with Table I. Rain/no rain contingency table of the WRF simulation
the Advanced Research WRF model, version 3.1. WRF is a against observation
nonhydrostatic, primitive-equation, mesoscale meteorological
model with advanced dynamics, physics and numerical WRF/observations Rain No rain
schemes. Detailed descriptions of the model can be found in
Rain RR (hits) RN (false alarms)
the model manual (Skamarock et al., 2008) and also on the No rain NR (misses) NN (correct negatives)
WRF user Web site (http://www.mmm.ucar.edu/wrf/users).

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
SENSITIVITY OF WRF MODEL IN RAINFALL SIMULATION 3015

through the following equations: total number of the time steps in the storm event). The final
values of RMSE, MBE and SD in the two dimensions are
1X N
RRi represented as percentages of the mean values of the
POD ¼ (1) corresponding observations. Because the random errors
N i¼1 RRi þ NRi
would be cancelled out with time during the rainfall-runoff
1X N
RRi þ RNi transformation, the bias index MBE, which is more
FBI ¼ (2)
N i¼1 RRi þ NRi representative of the error of the total rainfall amount, is
thus more crucial in hydrological applications. The other
1X N
RNi two indices of SD (representing the overall magnitude of
FAR ¼ (3)
N i¼1 RRi þ RNi random errors) and RMSE (where random errors could be
1X N
RRi accumulated during the calculation) are less important than
CSI ¼ (4) the MBE. Not a single verification index is perfect. The
N i¼1 RRi þ RNi þ NRi
characteristics and limitations of all the categorical and
continuous verification indices used in this study are
For the verification by categorical indices in the spatial summarised in Table II.
dimension, the results of the WRF model are first compared
with the observations of the rain gauges at each time step i,
and then the values of the categorical indices at all the time DATA AND EXPERIMENTAL DESIGN
steps are averaged to produce the final verification results.
Therefore, in Equations 1– 4, N refers to the total time steps The Brue catchment is chosen as the study area for the
of the simulation run. On the other hand, for the rainfall simulation using the WRF model. It is located in
verification in the temporal dimension, the indices are first southwest England (51.08 N and 2.58 W) with a drainage
calculated using the time series data of simulations and area of 135.2 km2. It is a predominantly rural catchment of
observations at each rain gauge i, then the averaged index modest relief with spring-fed headwaters rising in the
values of all the rain gauges are regarded as the final Mendip Hill and Salisbury Plain. The rain gauge network
verification results. Thus, instead of the simulation time consists of 49 Casella 0.2 mm tipping bucket type rain
steps, N represents the total number of the rain gauges in gauges. The observed rainfall data used in this study as the
Equations 1– 4 for temporal verification. ground truth of the simulated rainfall are obtained from the
The categorical indices can only provide a general Natural Environment Research Council–funded Hydro-
evaluation of the model performance on the basis of the logical Radar Experiment project, which ran from May
correctness of the simulated rainfall occurrences, whereas 1993 to April 1997 in the Brue catchment. Its data
for a more quantitative calculation of the simulation error, collection was extended to 2000.
continuous indices are needed. The term ‘continuous’ The five domain configuration scenarios, used in this
refers to the nature of the simulation. Although the WRF study, are designed to test the sensitivity of the rainfall
model generates only a discrete representation of continu- simulation of WRF using different downscaling ratios and
ous rainfall signals, the continuous assumption is reason- grid spacing. First, triple nested domains are centred over
able for computers with high enough machine precision the Brue catchment with approximately fixed domain sizes
(Jolliffe and Stephenson, 2003). The equations for of 900  900, 300  300 and 60  60 km2, respectively,
calculating the continuous indices RMSE, MBE and SD from the outermost domain (domain 1) to the innermost
are shown as follows: domain (domain 3). The resolution of the innermost
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi domain is fixed with a horizontal grid spacing of 1 km.
u M
u1 X 2
RMSE ¼ t Sj  Oj (5) Then four different downscaling ratios (1:10, 1:7, 1:5 and
M j¼1 1:3) are applied to increase the horizontal resolution from
the outermost domain to the innermost domain. In other
words, the grid spacing, which is fixed at 1 km in the
1X M  
MBE ¼ Sj  Oj (6) innermost domain, is increased by a fixed ratio for the outer
M j¼1 domains. The downscaling ratio here refers to the grid
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi spacing of the child domain divided by that of the mother
u
u 1 X M  2 domain. The downscaling ratio from domain 1 to domain 2
SD ¼ t Sj  Oj  MBE (7) and from domain 2 to domain 3 is kept the same. Taking
M  1 j¼1
the ratio of 1:10 as an example, with the grid spacing of
domain 3 fixed at 1 km, the grid spacing of domain 2 and
For spatial verification of the WRF model using domain 1 is increased to 10 and 100 km, respectively, by
continuous indices, Sj and Oj are the simulated and dividing the downscaling ratio of 1:10. Table III lists the
observed rainfall accumulations at each rain gauge j during variations of the grid spacing and the grid number with the
the whole storm event, (M in Equations 5–7 refers to the downscaling ratio changing from 1:10 to 1:3, which are
total number of the rain gauges), whereas for the temporal then regarded as the four domain configuration scenarios
verification, Sj and Oj are the simulated and observed from scenario 1 to scenario 4. It should be mentioned that
average areal rainfall at each time step j (M changes to the the sizes of the two outermost domains in scenario 2 and

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
3016 J. LIU, M. BRAY AND D. HAN

Table II. Ranges and characteristics of the categorical and continuous verification indices

Indices Range Perfect score Characteristics

Categorical POD 0–1 1 Indicates what fraction of the observed rains is


correctly simulated; ignores ‘false alarms’ and sensitive
to the frequency of rainfall occurrence during the event
FBI 0–1 1 Indicates whether WRF has a tendency to underestimate
(FBI < 1) or overestimate (FBI > 1) rainfall occurrences;
does not measure how well the simulation corresponds
to the observation
FAR 0–1 0 Indicates what fraction of the simulated rainfall does not
occur; ignores ‘misses’ and sensitive to the frequency
of rainfall occurrence during the event
CSI 0–1 1 Indicates how the simulated rainfall corresponds to the
observed rainfall; sensitive to ‘hits’, penalises both
‘misses’ and ‘false alarms’; does not distinguish
sources of simulation error
Continuous RMSE 0–1 0 Measures the average magnitude of error;
does not indicate the direction of deviations
and puts greater influence on large errors
MBE 1–1 0 Measures the average cumulative error including the
direction; does not measure the correspondence
between simulations and observations
SD 0–1 0 Measures the variation of the simulation error
about the MBE, which reflects the overall magnitude
of the random error; does not indicate the error direction

Table III. Configuration scenarios of triple/quadruple nested domains with different grid spacing and downscaling ratios

Scenario and domain Time step (h) Grid spacing (km) Grid number Domain size (km) Downscaling ratio

Scenario 1 (S1) Domain 1 3 100 99 900  900 –


Domain 2 1 10 30  30 300  300 1:10
Domain 3 0.25 1 60  60 60  60 1:10
Scenario 2 (S2) Domain 1 3 49 18  18 882  882 –
Domain 2 1 7 42  42 294  294 1:7
Domain 3 0.25 1 63  63 63  63 1:7
Scenario 3 (S3) Domain 1 3 25 36  36 900  900 –
Domain 2 1 5 60  60 300  300 1:5
Domain 3 0.25 1 60  60 60  60 1:5
Scenario 4 (S4) Domain 1 3 9 99  99 891  891 –
Domain 2 1 3 99  99 297  297 1:3
Domain 3 0.25 1 60  60 60  60 1:3
Scenario 5 (S5) Domain 1 3 27 55  55 1485  1485 –
Domain 2 3 9 99  99 891  891 1:3
Domain 3 1 3 99  99 297  297 1:3
Domain 4 0.25 1 60  60 60  60 1:3

The downscaling ratio refers to the grid size of the child domain divided by that of its mother domain, which varies among 1:10, 1:7, 1:5 and 1:3.

scenario 4 cannot be set at exact 900  900 and steps of the three domains, which also refer to the time
300  300 km2; however, they are made as close to them intervals of the output rainfall series, are set to 3 h, 1 h and
as possible (Table III). The different size of the outermost 15 min, respectively, from the outermost to the innermost
domain will result in slight difference of the IC and LBC domain and are kept the same for the four scenarios.
interpolated from the ERA-40 data in these two cases. It may be noticed that the grid spacing of domain 1 in
Because the difference of the domain size is less than 4%, it scenario 4 is 9 km and is far less than the resolution of the
is assumed that the variations of the downscaling results are original ERA-40 data, which is 1  1 and approximately
mainly caused by the downscaling ratios rather than the 100 km of the grid spacing. Because the IC and LBC of
differences of the boundary conditions caused by the the outermost domain are set by interpolating the global
slightly changed domain size. We can see from Table III data, a larger resolution gap between the global data and
that on the basis of the nearly fixed domain sizes and the the outermost domain means less accurate IC and LBC,
changing downscaling ratios, a certain domain could have which therefore has a negative impact on the downscaling
different grid numbers in different scenarios. The time results of the innermost domain. To further verify this,

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
SENSITIVITY OF WRF MODEL IN RAINFALL SIMULATION 3017

another scenario is designed as scenario 5 with quadruple Eight storm events with the same duration of 24 h are
nested domains for a comparison with scenario 4. By selected from the Brue catchments to run the WRF model
keeping the inner triple domains exactly the same as those with the five domain configuration scenarios. The durations
of scenario 4, we add another domain with a coarser of the eight events and the accumulative amounts of the
resolution and also a larger size as the outermost domain in average areal rainfall observed in the Brue catchment are
scenario 5 (see Table III). The grid spacing of this shown in Table IV. The areal rainfall is calculated by
outermost domain is 27 km, which is obtained following averaging the observations of the 49 gauges using the
the same downscaling ratio of 1:3. Thus, the inner domains Thiessen polygon method (Han and Bray, 2006). It should
of scenario 5 (domain 2, domain 3 and domain 4) can be be noted that in practice a warm-up period of 6 h is added
compared with the three domains in scenario 4 (domain 1, before the start of the storm event to run the WRF model.
domain 2 and domain 3), which have exactly the same grid The eight events are later categorised into four storm types
spacing and domain sizes, to see whether adding a bridge according to the evenness of the rainfall distribution in both
domain (domain 1 in scenario 5) can bring any benefit to space and time. In addition, the performances of the WRF
the downscaling results. Meanwhile, the outermost domain model in reproducing the four types of storm events are
of scenario 5 (domain 1) has similar grid spacing as that of compared based on the simulation results from the best
the outermost domain (domain 1) of scenario 3, so another scenario.
comparison can also be made to check whether a detailed
setting of the inner domains would make any difference to
the final results. The dimensions of the nested domains for RESULTS AND ANALYSES
the five scenarios concerning the location of the Brue
catchment are shown in Figure 1.
Sensitivity of the WRF model to different downscaling ratios
The WRF model is used to downscale the ERA-40
reanalysis data for eight 24-h storm events using the five
domain configuration scenarios of differing downscaling
ratios. For the hydrological use of the WRF rainfall
outputs, the accuracy of the total areal rainfall and its
temporal variation are important for lumped hydrological
models, whereas the spatial and the temporal rainfall
distributions are crucial factors when applied with
distributed models. Therefore, both the total rainfall
amount and the spatial and temporal distribution of the
WRF outputs are evaluated in this study. For a more
accurate calculation of the total rainfall amount, the WRF
areal rainfall is calculated by averaging the simulated
values at all the grids located inside the catchment
boundary rather than by averaging the point values
extracted at the rain gauge locations. The extracted
simulations at the rain gauges are only used for verification
index calculations. Table V lists the rankings of the five
scenarios on the basis of the accuracy of the WRF
simulated rainfall totals in the innermost domain. It is
Figure 1. Dimensions of the nested domains for the five scenarios that are noteworthy that most of the results have negative errors,
centred over the Brue catchment. Domain 1, domain 2 and domain 3 refer which indicate an underestimation of the total rainfall. One
to the three domains of S1–S4, which also represent domain 2, domain 3
and domain 4 of S5. The boundary shows the size of the outermost domain exception is event E, where all the five scenarios fail to
of S5 with an area of 1485  1485 km2 give any reasonable result with the simulation error

Table IV. Durations and accumulative rainfall amounts of the eight 24-h storm events

Event ID Storm start time Storm end time Accumulated 24-h rainfall (mm)

A 01/02/2000 04:00 01/02/2000 04:00 23.45


B 02/04/2000 18:00 03/04/2000 18:00 31.36
C 05/11/1999 06:00 06/11/1999 06:00 16.93
D 24/10/1999 00:00 25/10/1999 00:00 29.39
E 07/06/1996 11:00 08/06/1996 11:00 21.26
F 03/08/1994 13:00 04/08/1994 13:00 22.27
G 05/08/1997 10:00 06/08/1997 10:00 30.39
H 06/09/1995 18:00 07/09/1995 18:00 32.41

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
3018 J. LIU, M. BRAY AND D. HAN

Table V. Rankings of the five configuration scenarios for the simulation of the total rainfall amount in the innermost domain with 1-km
grid spacing

Event ID / Ranking (error%) First Second Third Fourth Fifth

A S3 (0.79) S2 (5.50) S1 (13.18) S4 (14.89) S5 (43.16)


B S3 (0.71) S4 (1.21) S1 (1.42) S5 (22.56) S2 (23.65)
C S4 (22.17) S3 (27.21) S5 (29.01) S2 (32.90) S1 (47.88)
D S4 (41.35) S1 (44.47) S2 (47.55) S5 (54.99) S3 (56.28)
E S5 (98.83) S2 (99.98) S1 (99.99) S3 (100.00) S4 (100.00)
F S2 (8.93) S3 (11.04) S5 (71.63) S4 (73.71) S1 (98.62)
G S4 (25.80) S3 (49.09) S5 (53.16) S2 (58.53) S1 (228.88)
H S2 (24.14) S4 (29.15) S5 (29.55) S3 (55.28) S1 (73.73)

Values in the parentheses represent the simulation errors as percentages of the observed amounts of the 24-h rainfall of each event.

reaching up to 100% because of the poor IC and LBC a single event (event A) is adopted to show the variance
provided by the ERA-40 data. A detailed analysis of the of WRF performance in different domains. The consistency
results of event E is presented in the Sensitivity of the WRF of the scenario ranking is also checked between the
Model to Different Storm Types Section. The remaining simulation of the rainfall amount and the rainfall
events have acceptable rainfall simulations for all five distribution in space and time. Because the outermost
scenarios, with the model errors of the best scenarios domains of all the scenarios have an output interval of 3 h
ranging from 0.71% (S3 for event B) to 41.35% (S4 for (which means the output files are written at 0, 3, 6, 9, 12,
event D). Overall, scenarios S2, S3 and S4 have similar 15, 18 and 21 Coordinated Universal Time (UTC) each
rankings. However, of these scenarios, S3 outperforms the day), for a feasible comparison of different domains,
other two, by ranking first place on two occasions (for simulation results with a duration of 27 h are used for the
event A and event B) and by having the lowest model analysis of event A, which starts at 01/02/2000 03:00 and
errors of 0.79% and 0.71%. By contrast, S1 is the worst ends at 01/02/2000 06:00. The categorical indices (POD,
scenario, ranked last place four times. S1 has the coarsest FBI, FAR and CSI) together with the continuous indices
downscaling ratio of 1:10. This downscaling ratio is steep (RMSE, MBE and SD) are calculated for the 27-h duration
(downscaling occurs from 100 to 1 km) and is the likely in both spatial and temporal dimensions. The values of the
cause for the poor performance of S1. With the exception two-dimensional indices are presented in Table VI and
of event E and event F, S5 is always ranked after S4 for the Table VII with respect to different domains and scenarios.
remaining events. This is an unexpected result because S5 Higher POD and CSI indices together with lower values
has one more outer domain than S4, and intuition dictates of FBI, FAR and the three continuous indices of RMSE,
that S5 should function better because of a more moderate MBE and SD indicate a better performance of the WRF
downscaling gradient. model. Considering the spatial verification indices in
For a detailed comparison of the five scenarios regarding Table VI, it can be seen that for all the scenarios, there is
the simulation of the rainfall distribution in space and time, an overall improvement from the outermost to the

Table VI. Spatial verification indices of event A for different domains of the five scenarios

Categorical indices Continuous indices

Scenario and domain POD FBI FAR CSI RMSE (%) MBE (%) SD (%)

Scenario 1 Domain 1 0.88 6.88 0.44 0.56 49.65 48.41 158.79


Domain 2 0.71 0.97 0.18 0.63 21.31 16.68 56.21
Domain 3 0.89 1.85 0.23 0.69 19.62 14.56 49.45
Scenario 2 Domain 1 0.88 6.38 0.44 0.51 38.55 36.94 121.40
Domain 2 0.71 0.83 0.18 0.60 12.61 0.35 12.79
Domain 3 0.91 2.69 0.24 0.69 13.16 3.93 18.07
Scenario 3 Domain 1 0.88 6.38 0.37 0.56 22.30 18.88 62.94
Domain 2 0.81 5.21 0.30 0.58 12.86 3.61 17.17
Domain 3 0.91 3.16 0.33 0.63 12.32 2.30 14.36
Scenario 4 Domain 1 0.75 3.50 0.28 0.56 20.94 17.22 57.61
Domain 2 0.76 3.02 0.23 0.61 19.33 15.56 52.23
Domain 3 0.90 3.32 0.29 0.66 19.00 15.25 51.21
Scenario 5 Domain 1 0.88 6.38 0.37 0.56 48.32 47.00 154.21
Domain 2 0.88 6.38 0.28 0.63 46.29 44.46 146.08
Domain 3 0.71 1.83 0.23 0.60 45.82 44.17 145.06
Domain 4 0.88 1.79 0.28 0.65 45.64 44.01 144.54

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
SENSITIVITY OF WRF MODEL IN RAINFALL SIMULATION 3019

Table VII. Temporal verification indices of event A for different domains of the five scenarios

Categorical indices Continuous indices

Scenario and domain POD FBI FAR CSI RMSE (%) MBE (%) SD (%)

Scenario 1 Domain 1 0.99 1.61 0.38 0.62 79.35 48.41 66.39


Domain 2 0.99 1.15 0.13 0.86 78.60 16.68 82.68
Domain 3 1.00 1.25 0.20 0.80 88.95 14.56 158.87
Scenario 2 Domain 1 0.99 1.77 0.44 0.56 79.72 36.94 74.53
Domain 2 0.99 1.15 0.13 0.86 92.31 0.35 94.00
Domain 3 1.00 1.30 0.23 0.77 101.33 3.93 107.80
Scenario 3 Domain 1 0.99 1.58 0.37 0.63 104.82 18.88 108.69
Domain 2 0.99 1.32 0.24 0.75 121.73 3.61 124.04
Domain 3 1.00 1.40 0.29 0.71 127.35 2.30 129.61
Scenario 4 Domain 1 0.99 1.34 0.26 0.73 127.71 17.22 133.40
Domain 2 0.99 1.22 0.19 0.81 133.73 15.56 137.55
Domain 3 0.99 1.34 0.26 0.74 140.34 15.25 197.04
Scenario 5 Domain 1 0.99 1.43 0.30 0.70 87.97 47.00 78.48
Domain 2 0.99 1.38 0.28 0.72 81.31 44.46 71.85
Domain 3 0.99 1.18 0.16 0.84 84.00 44.17 101.67
Domain 4 1.00 1.30 0.23 0.77 88.08 44.01 406.94

innermost domain with the increasing values of CSI and the temporal error of the WRF model in different domains
POD and decreasing values of FBI, FAR, RMSE, MBE and with different scenarios. Thus, with the increasing
and SD. For the comparison between different configur- values of RMSE and SD, we can conclude that an
ation scenarios, it should be mentioned that only the increasing resolution in space may not always guarantee
innermost domains (domain 4 for S5 and domain 3 for S1– better results in the temporal dimension because of the
S4) have the same horizontal resolution (1-km grid higher variability of the rainfall time series compared with
spacing), which is comparable among the five scenarios. the rainfall variability in space. By comparing the model
All the other domains of different scenarios are not performance of different scenarios with the innermost domain,
compared because they have different resolutions. By it is interesting to find that the best scenarios identified by
comparing the indices of the innermost domains, S3 is different indices are not identical. On the basis of the
found to be the best scenario, having the lowest values for categorical indices and RMSE, the best scenario is S1, whereas
all of the three continuous indices. Although the categorical on the basis of SD and MBE values, the best scenarios are S2
indices of S3 do not show their superiority over the other and S3, respectively.
scenarios, as we mentioned before, the categorical indices As we mentioned earlier, S5 is designed as a
can only measure the correctness of the rainfall occurrence complement to S4 to see whether there is any improvement
or nonoccurrence, which are less reliable thus not decisive of the model performance by using an additional domain to
compared with the continuous indices. provide a more gradual downscaling gradient from the
The temporal indices in Table VII reveal different ERA-40 data to the 1-km resolution. The three domains of
verification results of the WRF model in the temporal S4 (domain 1, domain 2 and domain 3) are compared,
dimension. In comparing the different domains, the four respectively, with the domains of S5 having the same
categorical indices together with MBE show an overall horizontal resolutions (domain 2, domain 3 and domain 4).
improvement of the model performance from the outermost With respect to the spatial indices, although the categorical
to the innermost domain. This is in line with the conclusion indices of domain 2 of S5 seem to be better than domain 1
made from the spatial verification. On the contrary, the of S4, all the categorical and continuous indices in the two
other two continuous indices, RMSE and SD, show an inner domains of S5 show no improvement at all. For the
increasing trend from the outermost to the innermost temporal verification, the categorical indices of S5 are
domain, indicating a deterioration of the downscaling found to be slightly better than those of S4, and the
performance in the temporal dimension. It might be noticed continuous indices RMSE and SD also show some kind of
that the MBE values in Table VII are exactly the same as improvement (in all the three domains shown by RMSE
those in Table VI. This is due to the characteristic of the and in the outer two domains of S4 by SD). However,
statistic itself. Calculations from two different dimensions considering the values of MBE in the three domains, which
can lead to the same results of MBE. Therefore, we can see is regarded as an overall estimation of the model error of
that the spatial and temporal aspects of the model bias error the two-dimensional rainfall distributions, generally there
cannot be separated using the MBE index. The values of is no obvious improvement in S5 compared with S4.
MBE in Table VI and Table VII should be interpreted as Moreover, another comparison can also be made between
an overall evaluation of the simulated rainfall distribution domain 1 of S5 and domain 1 of S3, which have similar
in the two dimensions of space and time. In this case, horizontal resolutions of 27 and 25 km, respectively, to
RMSE and SD are the only two reliable indices reflecting check whether a more detailed setting of the inner domains

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
3020 J. LIU, M. BRAY AND D. HAN

can make any difference to the downscaling. The results 30

turn out to be similar to the comparison between S4 and S5: (a) domain 1
there is no obvious improvement shown by the spatial indices. Observed

Cumulative rainfall (mm)


But for temporal verification, S5 is slightly improved with Scenario 1
respect to the categorical indices together with RMSE 20 Scenario 2
and SD. Scenario 3
Table VIII and Figure 2 provide another insight into the Scenario 4
comparison of the five scenarios regarding the simulated Scenario 5 dom2
rainfall accumulation in different domains. Table VIII 10
Scenario 5 dom1
presents the total amounts of the simulated rainfall and the
relevant model errors of event A in different domains
resulted from the five configuration scenarios, and Figure 2
illustrates the rainfall cumulative curves of the simulations
0
and observations in different domains. Again, for a feasible
comparison, the 27-h duration is used for event A. The 30
simulated areal rainfall in different domains is calculated by
averaging the WRF values at all the grids located inside the (b) domain 2
catchment boundary, whereas the observed areal values are

Cumulative rainfall (mm)


Observed
obtained by averaging the rain gauge observations using the Scenario 1
20
Thiessen polygon method. It can be seen from Table VIII Scenario 2
that the model error decreases for all the five scenarios from Scenario 3
the outermost to the innermost domain, which reveals an Scenario 4
improvement of the model performance in the simulation of Scenario 5 dom3
the total rainfall amount during the downscaling process. 10
For a detailed comparison, the rainfall accumulative
process is shown in Figure 2. Because the output data
intervals of both domain 1 and domain 2 of S5 are 3 h, the
curves of those two domains are shown in Figure 2a, 0
together with other 3-h output domains of S1–S4. Similarly,
the curves of domain 3 and domain 4 of S5 are presented in 30

Figures 2b and 2c, respectively. Figures 2a–2c demonstrate


(c) domain 3
that WRF simulated rainfall occurs at the same time as the
observed rainfall. For the cumulative curves of the Observed
Cumulative rainfall (mm)

innermost domain in Figure 2c, the best result is achieved Scenario 1


20
Scenario 2
with S3, which has almost the same accumulated rainfall as
Scenario 3
the observation by the end of the 27-h duration with 0.79%
Scenario 4
error (Table VIII). However, it should be mentioned that the
Scenario 5 dom4
difference in performance between S3 and S2 is not obvious
10
because a similar result is achieved by S2 in domain 2 with a
model error of 0.98% (Figure 2b and Table VIII). As we
concluded before, there is no obvious improvement in the
two-dimensional verification indices of S5 compared with
those of S4. This is confirmed by Table VIII and the 0

cumulative curves in Figure 2. S1, S2 and S3 show Figure 2. Cumulative curves of the simulated and observed catchment
areal rainfall of event A in different domains resulted from the five
improvement in the simulated rainfall with finer domain configuration scenarios
resolution; however, this is not the case with S4 and S5: no
significant difference is found between any of the might be the inappropriate domain setting of the two
cumulative curves of either S4 or S5 as shown from scenarios, which will be fully addressed in the Discussion
Table VIII and Figure 2. The most likely reason for this Section.
Table VIII. Total amounts of the catchment areal rainfall (mm) simulated in different domains of the five configuration scenarios for
event A with a 27-h duration

Scenario 1 (%) Scenario 2 (%) Scenario 3 (%) Scenario 4 (%) Scenario 5 (%)

Domain 1 12.14 (48.21) 14.92 (36.35) 19.27 (17.80) 19.54 (16.66) 12.489 (46.73)
Domain 2 19.79 (15.61) 23.68 (0.98) 22.90 (2.33) 19.90 (15.15) 13.324 (43.17)
Domain 3 20.36 (13.18) 24.74 (5.50) 23.26 (0.79) 19.96 (14.89) 13.318 (43.20)
Domain 4 – – – – 13.328 (43.16)

Values in the parentheses represent the model error as percentages of the observed value, which is 23.45 mm.

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
SENSITIVITY OF WRF MODEL IN RAINFALL SIMULATION 3021

Sensitivity of the WRF model to different storm types of the observed and the simulated catchment areal rainfall
are shown in Figure 11. Again, the observed areal rainfall is
The performance of the WRF model in simulating the eight calculated using the Thiessen polygon method whereas the
storm events is compared in this section to investigate the simulated values are obtained by averaging the grid values
sensitivity of the model to different storm types. Consider- inside the catchment boundary.
ing the better performance of S3 (resulting in the smallest The eight storm events are categorised into four main
errors of events A and B compared with the other events), types according to the evenness of the rainfall distribution
simulation results from scenario 3 are used for all the eight in the spatial and temporal dimensions. Although the storm
events for this comparison. The spatial distribution of the events can also be categorised from the meteorological
total amounts of rainfall occurring during the 24-h period of aspect, the hydrologists care more about the variations of
the eight events is shown in Figures 3–10, respectively, for the rainfall quantity in space and time, which have direct
events A–H. The observed rainfall accumulations at the impacts in hydrological applications. It should also be
rain gauges are shown in Figures 3a–10a, and the 1  1-km pointed out that the sensitivity of the WRF model is highly
grid-based simulated accumulations from the innermost related to the parameterisations, which might be suitable
domains are shown in Figures 3b–10b. It should be for one ‘storm type’ but inappropriate for another. Because it
mentioned that for some events, not all the 49 rain gauges is difficult to tell the best choice for future, the parameter-
worked during the 24-h storm, so only the valid rain gauges isations are normally fixed beforehand in operational
and their observed values are shown in the subfigures. To application. The following analyses of the WRF sensitivities
display the temporal characteristics of the rainfall for the are made based on the most commonly used parameterisa-
eight events, the cumulative curves and the time series bars tion schemes, which are fixed for different storm types.

(a) Observed (b) Simulated

Figure 3. Spatial distribution of (a) observed and (b) simulated rainfall accumulation over the catchment during storm event A (from 2000/02/01 04:00 to
2000/02/02 04:00)

(a) Observed (b) Simulated

Figure 4. Spatial distribution of (a) observed and (b) simulated rainfall accumulation over the catchment during storm event B (from 2000/04/02 18:00 to
2000/04/03 18:00)

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
3022 J. LIU, M. BRAY AND D. HAN

(a) Observed (b) Simulated

Figure 5. Spatial distribution of (a) observed and (b) simulated rainfall accumulation over the catchment during storm event C (from 1999/11/05 06:00 to
1999/11/06 06:00)

(a) Observed (b) Simulated

Figure 6. Spatial distribution of (a) observed and (b) simulated rainfall accumulation over the catchment during storm event D (from 1999/10/24 00:00 to
1999/10/25 00:00)

(a) Observed (b) Simulated

Figure 7. Spatial distribution of (a) observed and (b) simulated rainfall accumulation over the catchment during storm event E (from 1996/06/07 11:00 to
1996/06/08 11:00)

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
SENSITIVITY OF WRF MODEL IN RAINFALL SIMULATION 3023

(a) Observed (b) Simulated

Figure 8. Spatial distribution of (a) observed and (b) simulated rainfall accumulation over the catchment during storm event F (from 1994/08/03 13:00 to
1994/08/04 13:00)

(a) Observed (b) Simulated

Figure 9. Spatial distribution of (a) observed and (b) simulated rainfall accumulation over the catchment during storm event G (from 1997/08/05 10:00 to
1997/08/06 10:00)

(a) Observed (b) Simulated

Figure 10. Spatial distribution of (a) observed and (b) simulated rainfall accumulation over the catchment during storm event H (from 1995/09/06 18:00
to 1995/09/07 18:00)

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
3024 J. LIU, M. BRAY AND D. HAN

50 0 50 0

Rainfall intensity (mm/15min)


Rainfall intensity (mm/15min)
40 40
Cumulative rainfall (mm)

Cumulative rainfall (mm)


(a) Event a (b) Event b
1 1
30 Observed 30 Observed
Simulated Simulated

20 20
2 2

10 10

0 3 0 3

50 0 50 0

Rainfall intensity (mm/15min)


Rainfall intensity (mm/15min)
40 40

Cumulative rainfall (mm)


Cumulative rainfall (mm)

(c) Event c
1 1
30 Observed 30
Simulated (d) Event d
20 20 Observed
2 Simulated 2

10 10

0 3 0 3

100 0 100 0

Rainfall intensity (mm/15min)


Rainfall intensity (mm/15min)

80 80
Cumulative rainfall (mm)
Cumulative rainfall (mm)

2 2
(e) Event e (f) Event f
60 Observed 60 Observed
Simulated 4 Simulated 4

40 40

6 6
20 20

0 8 0 8

100 0 100 0
Rainfall intensity (mm/15min)
Rainfall intensity (mm/15min)

80 80
Cumulative rainfall (mm)

Cumulative rainfall (mm)

2 2
(g) Event g (h) Event h
60 Observed 60 Observed
Simulated Simulated 4
4
40 40

6 6
20 20

0 8 0 8

Figure 11. Time series bars and cumulative curves of the observed and simulated areal rainfall of the eight storm events

Events A and B have a relatively even distribution of the period. These two events can be classified into storm type
rainfall over the whole catchment (Figures 3a and 4a), and 1, which is characterised by an even distribution of rainfall
from the time series bars of Figures 11a and 11b, the in both spatial and temporal dimensions. It can be seen that
rainfall intensities of the two events are relatively even and the WRF model performs best in simulating and down-
the rainfall occurrence is continuous during a certain scaling this type of storms. The spatial distributions of

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
SENSITIVITY OF WRF MODEL IN RAINFALL SIMULATION 3025

rainfall are successfully reproduced (Figures 3b and 4b); of these two events by the WRF model are much better
the time series bars and the cumulative curves of the compared with those of type 3 events but are worse than
simulated results are quite close to the observations in the first two types. Actually, the classification of the four
Figures 11a and 11b. It should be noted that for event A in types of storm events could be made more quantitatively.
Figure 11a, the simulated bars start and end at exactly the Five widely applied indices (Van Etten, 2009; Bronikowski
same times as the observations. and Webb, 1996) are selected to measure the variability of
For the second type of storm events, the rainfall is the rainfall intensity in both spatial and temporal dimen-
distributed evenly in space, but its temporal distribution is sions. The equations of the indices are described as
uneven and discontinuous. The type 2 storms include event follows:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
C and event D among the eight events. The observed u N  2
u 1 X xi
rainfall of the two events has an even spatial distribution as Coefficient of variability ðCVÞ ¼ t 1 (8)
type 1 events (Figures 5a and 6a) and an uneven and N i¼1 x
discontinuous distribution in time (see the time series bars
in Figures 11c and 11d). For the type 2 storms, the rainfall 90P  10P
amounts could not be well simulated by the WRF model, Variability index ðVIÞ ¼ (9)
50P
but the spatial distributions (Figures 5b and 6b) are
reproduced evenly and the tendencies of the cumulative P
N
 pi lnðpi Þ
curves (in Figures 11c and 11d) are similar to those of the i¼1
observations, which indicate a good simulation of the Shannon index ¼ (10)
ln12
rainfall occurrence. X
N  
The observed rainfall of event E and event F has quite Simpson index ¼ 1  ln pi 2 (11)
distinctive characteristics. For the spatial distributions in i¼1
Figures 7a and 8a, the rainfall is highly concentrated in a !
small area, with the 24-h accumulation more than 40 mm, X
N
Berger  Parker index ¼ 1  x max  xi (12)
whereas other parts of the catchment have much less rain i¼1
with the least accumulation lower than 10 mm. With
respect to the time series bars in Figures 11e and 11f, it can
be observed that in both of the two events, the rainfall where xi represents the cumulative rainfall of each rain
durations are extremely short but quite intense. In this gauge i (for the variability of the spatial rainfall distribution)
case, we can categorise event E and event F into type 3 or the average areal rainfall at each modelling time step i (for
storms, the rainfall of which is densely concentrated in a the variability of the temporal rainfall distribution), x and
small area and occurs over a short time with large xmax are the mean and maximum values of xi and pi is the
intensities. The WRF model has the most difficulty in ratio of xi to the total amount of rainfall. For the calculation
reproducing this type of event. As aforementioned in the of VI in Equation 9, 90P, 10P and 50P are the 90th, 10th and
Sensitivity of the WRF Model to Different Downscaling 50th percentiles of the xi series.
Ratios Section, the model fails in capturing the whole The five variability indices are calculated in both spatial
process of the storm for event E (see Figures 7b and 11e); and temporal dimensions for the observed rainfall of the
for event F, although the total amount of the simulated eight storm events to check and further verify the
rainfall is close to the observations, the simulated process is classification of the four storm types we made before.
dislocated in the temporal dimension (see the time series For an easier comparison, the evenness of rainfall
bars in Figure 11f), and the concentrated areas in space are distribution of the four types of storm events is concluded
not identified correctly in Figure 8b. A direct inspection of in Table IX. The values of the five indices are shown in
the ERA-40 data helps find that the failure of the WRF Tables X and XI. Larger values of Shannon, Simpson
downscaling for this type of events is largely caused by the and Berger–Parker indices together with smaller values of
poor IC and LBC provided by the global data. A C-band CV and VI indicate more evenness of the rainfall
weather radar located at Wardon Hill was used to give a distribution. For the spatial evenness identified by the
full coverage of the Brue catchment during the Hydro- indices of CV and VI in Table X, we can see that type 1
logical Radar Experiment project (Moore et al., 2000).
Consistently, because of the highly concentrated processes
of this type of events, no rain was observed by the radar for Table IX. Evenness of the rainfall distribution for the four types
of storm events
event E and event F was also dislocated and under-
estimated. In space In time
The remaining two events, event G and event H, can be
regarded as type 4 events. This could be a mitigated storm Type 1 (events A and B) Y Y
type of type 3, which is also characterised by uneven and Type 2 (events C and D) Y N
discontinuous distributions of rainfall in space and time Type 3 (events E and F) N* N*
Type 4 (events G and H) N N
(referring to Figures 9a, 10a, 11g and 11h), but the
observed rainfall of this type is not highly concentrated and N*, unevenness with highly concentrated rainfall in a small area or a short
the intensity is much more moderate. The simulated results time.

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
3026 J. LIU, M. BRAY AND D. HAN

Table X. Spatial variability of the observed rainfall for the four types of storm events

Type 1 Type 2 Type 3 Type 4


Variability
indices Event A Event B Event C Event D Event E Event F Event G Event H

CV 0.11 0.14 0.10 0.09 0.37 0.28 0.13 0.14


VI 0.27 0.35 0.28 0.23 0.99 0.75 0.26 0.33
Shannon 1.00 1.00 1.00 1.00 0.98 0.99 1.00 1.00
Simpson 0.98 0.98 0.97 0.98 0.98 0.97 0.98 0.98
Berger–Parker 0.97 0.97 0.97 0.97 0.96 0.96 0.97 0.97

Table XI. Temporal variability of the observed rainfall for the four types of storm events

Type 1 Type 2 Type 3 Type 4


Variability
indices Event A Event B Event C Event D Event E Event F Event G Event H

CV 1.12 0.45 1.84 1.22 2.65 3.74 1.33 1.26


VI 5.15 1.28 15.27 5.84 125.56 9.22 7.65 7.35
Shannon 0.78 0.97 0.66 0.78 0.43 0.35 0.76 0.79
Simpson 0.91 0.95 0.82 0.90 0.68 0.40 0.89 0.90
Berger–Parker 0.86 0.92 0.65 0.81 0.56 0.23 0.82 0.81

and type 2 events have the lowest values, which represent performs moderately for type 2 and type 4 events, with
the best evenness of the spatial rainfall; type 4 shows type 2 slightly better than type 4. The simulation errors of
higher CVs and VIs compared with type 1 and type 2; and the total rainfall amounts for the four types of events are
the two events of type 3 are found to have the largest represented in Table XIV. An examination of the errors in
values of the two indices, which are consistent with the the three domains also leads to the same conclusions of the
most highly concentrated spatial rainfall we observed from model performance in simulating the different types of
Figures 7a and 8a. However, the three indices of Shannon, storm events. In general, the model performance can be
Simpson and Berger–Parker fail in making any distinction summarised by the ranking: type 1 > type 2 > type 4 > type
among the four storm types, with all the four types of 3, from the best to the worst.
events having similar values. With respect to the
identification of the temporal rainfall variability in
Table XI, all the five indices are found to be efficient.
DISCUSSION
Again, type 1 events have the lowest values of CV and VI
and the highest values of Shannon, Simpson and Berger– A two-part sensitivity study is carried out on the
Parker indices, which tell a most even rainfall distribution performance of the WRF model to different domain
in time. The two events of type 3 have the largest CV and settings and various storm types. The first part of the
VI values, accompanied by the lowest values of the other analysis assesses the five model configuration labelled
three indices, which indicate the most uneven temporal scenarios, whereas the second part of analysis concentrates
distribution. In addition, the five indices of type 2 and type on the ability of the WRF model to simulate various storm
4 reveal a case between type 1 and type 3, which is also in types with different rainfall distributions in space and time.
line with the moderate temporal unevenness of the two In both of the sensitivity studies, model performances are
types as we concluded from the time series bars and the evaluated using the two-dimensional verification indices as
cumulative curves in Figure 11. well as the rainfall totals.
The performances of the WRF model on the four types In the first part of the analysis, an examination on the
of storm events are further verified by using the two- total rainfall amounts simulated in the innermost domains
dimensional categorical and continuous indices. The of the five scenarios (with the same objective resolutions of
indices are calculated in the innermost domain of the eight 1 km) shows that the best model performances are achieved
events for both spatial and temporal dimensions, the values with scenario S2, S3 and S4. S1 turns out to be the worst
of which are shown in Tables XII and XIII. The results are scenario because it is consistently ranked last. Next, the
in agreement with our previous conclusions on the model five scenarios are compared more quantitatively by using
performance. The WRF model shows the best perfor- the two-dimensional verification indices calculated on the
mances for type 1 events in both space and time, with the basis of the simulation results of event A, with which the
highest values of POD and CSI and the lowest values of WRF model functions the best. All the spatial verification
FBI, FAR, RMSE, MBE and SD. Type 3 events have the indices show that the best results are obtained with S3,
worst index values among the four types. The model which is consistent with the conclusions based on the total

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
SENSITIVITY OF WRF MODEL IN RAINFALL SIMULATION 3027

Table XII. Spatial verification indices in the innermost domain for the four types of storm events by configuring the WRF model using
scenario 3

Categorical indices Continuous indices

Types of storm events POD FBI FAR CSI RMSE (%) MBE (%) SD (%)

Type 1 Event A 0.91 3.16 0.33 0.63 12.32 2.30 14.36


Event B 1.00 1.31 0.16 0.84 14.10 0.04 14.27
Type 2 Event C 0.61 3.01 0.45 0.38 33.79 30.46 148.53
Event D 0.65 2.32 0.32 0.47 57.67 56.95 138.34
Type 3 Event E 0.00 0.00  0.00 106.65 100.00 351.75
Event F 0.21 3.08 0.91 0.02 32.12 12.32 51.22
Type 4 Event G 0.87 6.47 0.56 0.37 53.37 48.18 111.98
Event H 0.51 2.86 0.51 0.29 59.39 56.76 118.10

Table XIII. Temporal verification indices in the innermost domain for the four types of storm events by configuring the WRF model
using scenario 3

Categorical indices Continuous indices

Types of storm events POD FBI FAR CSI RMSE (%) MBE (%) SD (%)

Type 1 Event A 1.00 1.40 0.29 0.71 120.14 2.30 122.12


Event B 1.00 1.20 0.16 0.84 78.64 0.04 79.05
Type 2 Event C 0.80 1.38 0.42 0.51 238.60 30.46 464.83
Event D 0.87 1.19 0.27 0.66 117.18 56.95 415.40
Type 3 Event E 0.00 0.00 – 0.00 375.54 100.00 901.44
Event F 0.06 1.40 0.91 0.03 533.91 12.32 549.25
Type 4 Event G 0.81 1.93 0.58 0.39 358.94 48.18 451.89
Event H 0.62 1.22 0.48 0.39 342.90 56.76 476.08

Table XIV. Simulation errors of the total rainfall amounts for the four types of storm events using scenario 3

Type 1 Type 2 Type 3 Type 3

Event A Event B Event C Event D Event E Event F Event G Event H

Domain 1 17.80 10.82 21.86 49.68 100.00 34.47 76.34 42.65


Domain 2 2.33 2.46 29.19 55.29 100.00 12.74 29.05 56.47
Domain 3 0.79 0.71 27.21 56.28 100.00 11.76 19.97 55.28

Values are presented in the percentages of the corresponding observations (%). Because of the inconsistency of the result output time of domain 1 (based
on a 3-h interval) and the event starting time, events A, E, F and G are calculated for a duration of 27 h, whereas the other four events of B, C, D and H
are calculated for a duration of 24 h.

rainfall amount of event A. Although for the temporal As stated, a coarse downscaling ratio such as 1:10 will lead
indices different indices give various results, besides S3, to large resolution gaps between the mother and the child
which is identified to be the best by the categorical domains in the nested system and is inappropriate for
temporal indices and MBE, S2 and S1 are also found to be downscaling data of resolutions in the order of 1  1 (e.g.
the best scenario regarding the values of SD and RMSE, ERA-40). The commonly used downscaling ratio of 1:3
respectively. However, because the index of MBE can (used in S4) does not always perform the best for all the
reflect an overall evaluation of the simulated rainfall cases, that is to say, other moderate ratios are also worth
distribution in both spatial and temporal dimensions, the trying when setting up the nested domains. There may also
least MBE value of S3 indicates it is the best scenario for be a lack of consistency, for a given scenario, in the
event A with respect to not only the total amount but also performance of WRF in simulating the total rainfall
the two-dimensional distributions. By checking the down- amount and the distribution in both spatial and temporal
scaling ratios of the five scenarios, it can be observed that dimensions. The choice of the downscaling ratios should
S2, S3 and S4 have moderate ratios of 1:7, 1:5 and 1:3, depend on the research purpose and also the characteristics
respectively, whereas S1 has the least refined ratio of 1:10. of the studied storm event.

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
3028 J. LIU, M. BRAY AND D. HAN

Another issue deserving explanation is the unsatisfactory decisive index for the overall evaluation of the simulated
results of S5. Better model performance using S5 had been rainfall distribution, it can be concluded that the model
expected because on the basis of the nested triple domains performance has been improved as we downscale the ERA-
in S4, adding another domain would reduce the resolution 40 data to the catchment scale. In addition to event A, an
gap between the ERA-40 data and the outermost domain of inspection of the other events can also lead to the same
S4. However, the actual performance of S5 does not show findings, although the results are not shown in this article.
any obvious improvement in either the total rainfall amount This is consistent with the conclusions in other studies we
or the two-dimensional distributions (albeit some temporal mentioned in the Introduction. However, it should be noted
indices show slightly better results compared with S4 for that this improvement is more likely to happen in the
event A). A possible reason might be that the added simulation of the rainfall amount and its spatial distribu-
domain has a larger size of approximate 1500  1500 km2, tion, not necessarily in the temporal dimension (i.e. the
which is nearly two thirds more than the size of the simulation of the temporal distribution of rainfall may not
outermost domain in S4. The increased computing area be improved) because of the higher variability of the
lays a heavy burden of further data processing and more rainfall time series compared with its spatial distribution.
uncertainties are in involved in the downscaling process. Moreover, as we stated before, this kind of improvement in
Moreover, domain 1 of S4 might get better setting by direct S4 and S5 is not as obvious as S1–S3, with respect to the
interpolating from the ERA-40 data than domain 2 of S5, differences of the nested domains in the simulated rainfall
which has the same size and resolution but is set by using amount and the two-dimensional indices. The same case
the information from its mother domain. To further prove also happens with the new scenario S6, which has the best
this, a new scenario is created (thereafter called S6) to performance among all the scenarios: there is also no
reveal the significance of the domain size. Four nested improvement observed from the 3- to the 1-km domain.
domains are used with the downscaling ratio set to 1:3. The All the three configuration scenarios of S4, S5 and S6
grid spacing of the domains from outermost to innermost is have the downscaling ratio of 1:3. The reason might be
27, 9, 3 and 1 km. However, in contrast to S5, the domain related to the quality of the ERA-40 data. The simulation
sizes are set to be 900  900, 300  300, 60  60 and in the 3-km domain might already achieve the best results
27  27 km2. That is to say, the sizes of the outer three that the ERA-40 data can afford, thus leaving no room for
domains of S6 are the same as the triple domains in S4. For further improvement made at finer resolutions. That is to
the new scenario, no further data will be processed (with no say, the effect of the downscaling is dependent on both the
increased size of the outermost domain) compared with S4, domain configurations and the quality of the IC and LBC
but only an added domain with a ‘bridge’ resolution that provided by the global data. This also indicates that
between the ERA-40 data and the 1-km objective with the size and the downscaling ratio of the nested
resolution. The WRF model is run again for the eight domains appropriately set, there is no need to downscale the
storm events following the domain settings in S6. By global data to a very fine resolution so that the modelling
comparing the simulation errors of the total rainfall time could be greatly shortened.
amounts in the innermost domains of all the scenarios, For the second part of sensitivity analysis conducted on
S6 turns out to be the best in simulating all the eight events. the WRF simulations of different storm events, the eight
In addition, even the third domain of S6 (with the grid events are first categorised into four main types according
spacing of 3 km) can have better results than the 1-km to the evenness of the spatial and temporal distributions of
domains of other scenarios. Now it is clear that only with the rainfall observations. The model simulations of the
the unchanged size of the outermost domain (i.e. the fixed eight events using scenario 3 are compared by examining
IC and LBC from the global data) can an added ‘bridge the spatial distributions and the cumulative curves together
domain’ improve the downscaling results. This is the with the time series bars of the simulated rainfall. The two
reason S4 functions better than S5, but worse than S6. type 1 events have a two-dimensional evenness with
Also, it can be seen that using the four nested domains will rainfall distributed evenly in space and continuously in
not always guarantee better results than using triple time, and it is found that WRF performs best in simulating
domains unless they are appropriately set with reasonable this type of storms. Type 2 events have only one-
domain sizes. dimensional evenness in space, whereas the spatial
Besides the comparison of different domain configu- distribution of rainfall can still be simulated evenly with
ration scenarios, the performances of the WRF model are the temporal trends found to be consistent with the
also compared at different resolutions for event A. Again, observations, but the total amounts are underestimated
both the rainfall amounts, and the spatial and temporal seriously. The two-dimensional unevenness is found with
distributions are used to evaluate the model performance in both type 3 and type 4 events. The WRF model did not
different domains. For event A, all the scenarios are found give satisfactory simulation results in either the spatial
to have the same trends: the total rainfall amounts and all distribution of rainfall or its temporal occurrence. For type
the spatial indices together with the categorical temporal 3 events, the unevenness is found to be much more serious
indices reveal an improvement of the model performance with the rainfall highly concentrated in small areas and
from the outermost to the innermost domain. However, the short times with large intensities. The rainfall is either
continuous temporal indices of RMSE and SD increase missed or serious dislocated by WRF because of the poor
with the increase of resolution. Because the MBE is the IC/IBCs from the ERA-40 data. Although the simulation

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
SENSITIVITY OF WRF MODEL IN RAINFALL SIMULATION 3029

results of type 4 are better than type 3, the total amounts of concept. Because of the limited storm events studied in this
rainfall are overestimated for both of the two events. The article, the definition of the absolute boundary between
classification of the four storm types is verified by five evenness and unevenness could not be given here. Further
commonly used variability indices, and the model studies on more storm events are needed.
performances of the four types of events are further Model verification has long been recognised as a tricky
checked by the two-dimensional verification indices. The issue in NWP modelling. In this study, a two-dimensional
classification of the storm types can help to find some verification scheme is proposed to measure the simulated
general patterns of the performance of the WRF model, for rainfall distribution in both spatial and temporal dimen-
example, it is interesting to find that type 2 events tend to sions, together with the amount of the total rainfall to
underestimate the total rainfall amount whereas type 4 is constitute an overall evaluation for the performance of the
usually overestimated. In that case, an error correction WRF model. A point that should be clarified is the
model could be built to improve the model performances assumption of the observed rainfall data as the ‘ground
on the basis of the general patterns of the model error with truth’ for the model verification. In contrast to ‘all models
different storm types. It is noteworthy that the conclusions are wrong, but some are useful’ (Box and Jenkins, 1970), it
are subject to the parameterisation schemes used in this is necessary to consider ‘all data are wrong, but some are
study and also to the classification method of the storm useful’ (Han, 2010). The observed rainfall data used in this
types. Future work should consider other parameterisation study are measured by the 49 rain gauges in the Brue
schemes together with storm events of various spatial and catchment, which are point-based measurements instead of
temporal rainfall distributions. Investigations also need to spatial observations. In addition to the instrumental
be carried out in other catchments with different geograph- measurement errors, the spatial representation of the point
ic and climatic conditions so that more general and useful values could introduce a challenging problem when they
patterns can be found. are used to assess the grid based spatial data (Ishak et al.,
It is useful to have some synoptic analysis for the 2010; Bray et al., 2010). In this study, we assume that the
meteorological characteristics of the storm events. How- rain gauges are reliable and the point measurements can
ever, in this study, all the eight events are found to happen represent the area of a single grid in each domain. That is
under the control of the cyclonic weather system and most why in the calculation of the two-dimensional indices, the
of the summer events are thunderstorms (according to the point observations are compared with the grid simulations
‘weather log’ of Royal Meteorological Society updated extracted at the locations of the 49 rain gauges. This is
monthly and maintained by Philip Eden). Besides events E currently the best we can do with the point-based
and F, which are easily identifiable as showers caused by observations from the rain gauges. If more observations
strong local convections, there is no big difference found are available, such as the weather radar data, the ‘ground
for the other events with respect to the climatic ingredients. truth’ would be much closer to the true values. It should be
More detailed differences might be hidden in the storm mentioned that the two-dimensional indices can also be
formation processes, which are difficult to detect by transferable to the verifications using grid-based observa-
traditional weather observing techniques. Moreover, be- tions by simply replacing the concept of rain gauges in
cause there are large uncertainties in the rainfall generation Equations 1–7 with grids that used for comparison with a
and development, the chances of a rainfall event having the reasonable size. Overall, this is an initial trial for the
same climatic characteristics in another catchment with verification of the NWP model in two dimensions.
similar geographic conditions are slim. Therefore, the Although a sophisticated facility named Model Evaluation
classification of the storm events regarding the evenness of Tools (MET), which is designed especially for the WRF
the spatial and temporal distribution seems to be rather model, could be used, the two-dimensional verification
useful in practice. As aforementioned, with clearer patterns scheme proposed in this study provides a simple way that
to be found with different types of storm events, error is easily applied for the verification of the NWP models. As
correction models can be developed to improve the a limitation of the application of the indices in this study,
downscaled results. Moreover, although the classification when comparing the verification results in different
is a posteriori, it can also help hydrologists in flow domains, it is better to convert the simulations into the
forecasting, that is, they may want to know the confidence same time intervals before calculating the indices.
of the downscaled rainfall products before being trans- Finally, it may be noticed that being one of the most
formed into flow through the hydrological models. The five challenging variables handled in NWP modelling, the
variability indices in this study are found to work simulated rainfall has considerably large errors (except type
effectively in providing a more quantitative evaluation of 1 events), which cannot be used directly for water
the evenness of rainfall distribution in either spatial or management or flood forecasting. However, in practice, the
temporal dimensions, which could be used in the future for data assimilation system is normally involved to reduce
a more comprehensive investigation of the WRF model and the errors. The WRF Data Assimilation (WRFDA) is a
further refinements in domain configurations. However, it variational data assimilation system used to ingest various
should be noted that the variability indices can be only sources of observations into the WRF products (the first guess
used for a comparison between different storm events or background forecast) and their respective error statistics to
rather than the identification of evenness. The evenness or provide an improved estimate of the atmospheric state, thus
unevenness mentioned in this study is some kind of relative helping to result in better performance of the WRF model

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
3030 J. LIU, M. BRAY AND D. HAN

(Skamarock et al., 2008). Weather radar is an important data configuration ratios indeed reflect the model’s sensitivity to
source for mesoscale weather analysis and forecasting (Xiao boundary conditions as a whole. Although the boundary
et al., 2005), and nowadays it has been proved by many condition effect has been fully discussed by the meteorol-
studies that the assimilation of the weather radar data could ogists (e.g. Nutter et al., 2004; Termonia et al., 2009; Vié
largely improve the performance of the WRF model in et al., 2011), this article addresses this issue for the purpose
rainfall modelling (Sugimoto et al., 2005; Xiao and Sun, of improving the downscaled rainfall products for hydro-
2007; Liu et al., 2008; Huang et al., 2009). Moreover, when logical applications, especially from the aspect of rainfall
using the rainfall products for flood forecasting, the real-time quantity variation in spatial and temporal dimensions.
updating scheme in the forecasting system could also help to However, it should be noted that the results and
further improve the final results. On the other hand, the large conclusions in this article are subject to the specific
errors of the modelled rainfall might also be caused by an domain setup method used for the configuration scenario
inappropriate choice of the cumulus parameterisation design and most importantly the particular set of
schemes (Mazarakis et al., 2009; Xuan et al., 2009), which parameterisation schemes adopted in this study. It is
is another sensitivity issue in NWP modelling of the rainfall. important that similar sensitivity studies are carried out
In this study, to better investigate the sensitivity of the model with other parameterisation schemes and in different
performance to different domain settings and storm types, we catchments with various geographic and climatic condi-
focused on the Kain–Fritsch scheme (Kain, 2004), which is tions so that the accumulated experience and knowledge
the most widely applied scheme in current studies. However, will help hydrologists to explore more general patterns of
other cumulus parameterisation schemes are worth trying to the model error and to further improve the accuracy of the
see whether the same conclusions can be found regarding the rainfall products from the mesoscale NWP models.
domains settings and the WRF performances with different
storm types.
ACKNOWLEDGEMENT

The authors thank the anonymous reviewers for their


CONCLUSIONS
insightful comments and valuable suggestions in improv-
In this study, the ECMWF ERA-40 data are used to drive ing the rigor and depth of the manuscript.
the WRF model for rainfall simulation of eight selected
storm events with durations of 24 h. Five domain
configuration scenarios are designed with different down- REFERENCES
scaling ratios to test the sensitivity of the WRF model to
Betts AK, Ball JH, Viterbo P, Dai A, Marengo J. 2005. Hydrometeorology
domain settings. The sensitivity of the model performance of the Amazon in ERA-40. Journal of Hydrometeorology 6: 764–774.
in simulating different storm types is also investigated by Box GEP, Jenkins GM. 1970. Time series analysis forecasting and
categorising the eight events into four storm types control. Holden Day: San Francisco; 553.
Brath A, Burlando P, Rosso R. 1988. Sensitivity analysis of real-time
according to the rainfall distribution in space and time. A flood forecasting to on-line rainfall predictions. In Selected Papers from
two-dimensional verification scheme is proposed and the Workshop on Natural Disasters in European-Mediterranean
found to be effective in evaluating the simulated rainfall Countries. Siccardi F, Bras RL (Eds). Perugia: Italy; 469–488.
Bray M, Han D, Xuan Y, Bates P, William M. 2010. Rainfall uncertainty
distributions in both spatial and temporal dimensions. The from NWP downscaling model. Hydrological Processes. DOI:10.1002/
downscaling ratio of 1:3, which has been recommended for hyp.7905.
the MM5 model and still widely used for WRF, is not Bronikowski A, Webb C. 1996. Appendix: A critical examination of
rainfall variability measures used in behavioral ecology studies.
always the best choice in dealing with different storm Behavioral Ecology and Sociobiology 39(1): 27–30. DOI: 10.1007/
events. Except for the downscaling ratio 1:10, which is s002650050263
found to be too coarse for the downscaling, other ratios Chang HI, Kumar A, Niyogi D, Mohanty UC, Chen F, Dudhia J. 2009.
The role of land surface processes on the mesoscale simulation of the
such as 1:7 and 1:5 can also result in good performance of July 26, 2005 heavy rain event over Mumbai, India. Global and
the WRF model. Besides the downscaling ratio, domain Planetary Change 67: 87–103.
size is another significant issue affecting the model Chen CS, Lin YL, Peng WC, Liu CL. 2010. Investigation of a heavy
rainfall event over southwestern Taiwan associated with a subsynoptic
performance. With appropriate settings of the downscaling cyclone during the 2003 Mei-Yu season. Atmospheric Research 95:
ratio and domain size, characterised by an overall 235–254.
improvement of the model performance from the outermost Cluckie ID, Han D. 2000. Fluvial Flood Forecasting. Water and
Environmental Management 14(4): 270–276.
to the innermost domain, the modelling time can be greatly Colle BA, Mass CF. 2000. The 5–9 February 1996 flooding event over the
reduced (e.g. for a comparison, the recorded modelling Pacific Northwest: sensitivity studies and evaluation of the MM5
times for event A with S4, S5 and S6 are 1330, 1274 and precipitation forecasts. Monthly Weather Review 128(3): 593–617.
Davis C, Brown B, Bullock R. 2006. Object-based verification of
201 min, respectively). For the sensitivity study on precipitation forecasts, Part I: methodology and application to
different storm types, the WRF model shows the best mesoscale rain areas. Monthly Weather Review 134: 1772–1784.
performance in reproducing the storms with the two- Done J, Davis CA, Weisman M. 2004. The next generation of NWP:
explicit forecasts of convection using the Weather Research and
dimensional evenness of rainfall distributions, whereas the Forecasting (WRF) model. Atmospheric Science Letter 5: 110–117.
storms with highly concentrated heavy rainfall in space and Dudhia J. 1989. Numerical study of convection observed during the winter
time are the trickiest case for the model to handle and are monsoon experiment using a mesoscale two-dimensional model.
Journal of the Atmospheric Sciences 46(20): 3077–3107.
likely to result in a failure in capturing the whole process Etherton, B, Santos B. 2008. Sensitivity of WRF forecasts for South
of the storm. The results of applying different domain Florida to initial conditions. Weather and Forecasting, 23, 725–740.

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)
SENSITIVITY OF WRF MODEL IN RAINFALL SIMULATION 3031

Fowle MA, Roebber PJ. 2003. Short-range (0-48 h) numerical prediction Moore RJ, Jones DA, Cox DR, Isham VS. 2000. Design of the HYREX
of convective occurrence, mode, and location. Weather and Forecasting raingauge network. Hydrology and Earth System Sciences 4(4):
18: 782–794. 523–530.
Fritsch JM, Carbone RE. 2004. Improving quantitative precipitation Nutter P, Stensrud D, Xue M. 2004. Effects of coarsely resolved and
forecasts in the warm season: a USWRP research and development temporally interpolated lateral boundary conditions on the dispersion
strategy. Bulletin of the American Meteorological Society 85(7): 955–965. of limited-area ensemble forecasts. Monthly Weather Review 132:
Graversen RG, Kallen E, Tjernstrom M, Kornich H. 2007. Atmospheric 2358–2377.
mass-transport inconsistencies in the ERA-40 reanalysis. Quarterly Shem W, Shepherd M. 2009. On the impact of urbanization on
Journal of the Royal Meteorological Society 133: 673–680. summertime thunderstorms in Atlanta: two numerical model case
Han D. 2010. Flood risk assessment and management. Bentham Science studies. Atmospheric Research 92: 172–189.
Publishers (in press): Bussum, the Netherlands, ISBN: 978-1-60805- Skamarock WC, Klemp JB, Dudhia J, Gill DO, Barker DM, Duda MG,
047-5 Huang XY, Wang W, Powers JG. 2008. A Description of the Advanced
Han D, Bray M. 2006. Automated Thiessen Polygon Generation. Water Research WRF Version 3. NCAR Technical Note, NCAR/TN-
Resources Research 42: W11502. DOI:10.1029/2005WR004365 475 + STR.
Hong SY, Lee JW. 2009. Assessment of the WRF model in reproducing a Stanski HR, Wilson LJ, Burrows WR. 1989. Survey of common
flash-flood heavy rainfall event over Korea. Atmospheric Research 93: verification methods in meteorology. World Weather Watch Technical
818–831. Report No. 8, WMO/TD No.358, WMO, Geneva, 114.
Hong SY, Dudhia J, Chen SH. 2004. A revised approach to ice Sugimoto S, Crook NA, Sun J, Barker DM, Xiao Q. 2005. Potential Benefits
microphysical processes for the bulk parameterization of clouds and of Multiple-Doppler Radar to Quantitative Precipitation Forecasting:
precipitation. Monthly Weather Review 132(1): 103–120. Assimilation of Simulated Data Using WRF-3DVAR System. In:
Hong SY, Noh Y, Dudhia J. 2006. A new vertical diffusion package with Proceedings of World Weather Research Program Symposium on
an explicit treatment of entrainment processes. Monthly Weather Review Nowcasting and Very Short Range Forecasting, Toulouse, France.
134: 2318–2341. Termonia P, Deckmyn A, Hamdi R. 2009. Study of the lateral boundary
Huang X, Xiao Q, Barker DM, Zhang X, Michalakes J, Huang W, condition temporal resolution problem and a proposed solution by
Henderson T, Bray J, Chen Y, Ma Z, Dudhia J, Guo Y, Zhang X, Won means of boundary error restarts. Monthly Weather Review 137:
D, Lin H, Kuo Y. 2009. Four-dimensional variational data assimilation 3551–3566.
for WRF: Formulation and preliminary results. Monthly Weather Toth E, Brath A, Montanari A. 2000. Comparison of short-term rainfall
Review 137: 299–314. prediction models for real-time flood forecasting. Journal of Hydrology
Ishak AM, Bray M, Remesan R, Han D. 2010. Estimating reference 239: 132–147.
evapotranspiration using numerical weather modelling. Hydrological Van Etten EJB. 2009. Inter-annual Rainfall Variability of Arid Australia:
Processes. DOI:10.1002/hyp.7770 greater than elsewhere? Australian Geographer 40(1): 109–120. DOI:
Jolliffe IT, Stephenson DB. 2003. Forecast Verification: A Practitioner’s 10.1080/00049180802657075
Guide in Atmospheric Science. John Wiley and Sons: Chichester, UK; Vié B, Nuissier O, Ducrocq V. 2011. Cloud-resolving ensemble
240. simulations of Mediterranean heavy precipitating events: uncertainty
Kain JS. 2004. The Kain–Fritsch convective parameterization: An update. on initial conditions and lateral boundary conditions. Monthly Weather
Journal of Applied Meteorology 43(1): 170–181. Review 139: 403–423.
Kain JS, Weiss SJ, Levit JJ, Baldwin ME, Bright DR. 2006. Examination Wang S, Huang S, Li Y. 2006. Sensitive numerical simulation and
of convection-allowing configurations of the WRF model for the analysis of rainstorm using nested WRF model. Journal of
prediction of severe convective weather: the SPC/NSSL spring program Hydrodynamics, Ser.B 18(5): 578–586.
2004. Weather and Forecasting 21: 167–181. Wilks DS. 2006. Statistical Methods in Atmospheric Science, 2nd Edn.
Knievel JC, Ahijevych DA, Manning KW. 2004. Using temporal modes Academic Press: Burlington, MA, 627.
of rainfall to evaluate the performance of a numerical weather Xiao Q, Sun J. 2007. Multiple radar data assimilation and short-range
prediction model. Monthly Weather Review 132: 2995–3009. Quantitative Precipitation Forecasting of a squall line observed during
Liu C, Xiao Q, Wang B. 2008. An ensemble-based four-dimensional IHOP 2002. Monthly Weather Review 135: 3381–3404.
variational data assimilation scheme: Part I: Technical formulation and Xiao Q, Kuo Y, Sun J, Lee W, Lim E, Guo Y, Barker DM. 2005.
preliminary test. Monthly Weather Review 136: 3363–3373. Assimilation of Doppler radar observations with a regional 3D-VAR
Mazarakis N, Kotroni V, Lagouvardos K, Argiriou AA. 2009. The system: impact of Doppler velocities on forecasts of a heavy rainfall
sensitivity of numerical forecasts to convective parameterization during case. Journal of Applied Meteorology 44: 768–788.
the warm period and the use of lighting data as an indicator for Xiu A, Pleim JE. 2001. Development of a Land Surface Model, Part I:
convective occurrence. Atmospheric Research 94: 704–714. Application in a Mesoscale Meteorological Model. Journal of Applied
Mlawer EJ, Taubman SJ, Brown PD, Iacono MJ, Clough SA. 1997. Meteorology 40(2): 192–209.
Radiative transfer for inhomogeneous atmosphere: RRTM, a validated Xuan Y, Ian D, Han D. 2005. Uncertainties in application of NWP-based
correlated-k model for the longwave. Journal of Geographical QPF in real-time flood forecasting. In: Proceedings of ACTIF
Research 102(D14): 16663–16682. International conference on innovation advances and implementation
Mölders N, Kramm G. 2010. A case study on wintertime inversions in of flood forecasting technology, Tromsø, Norway.
Interior Alaska with WRF. Atmospheric Research 95(2–3): 314–332. Xuan Y, Cluckie ID, Wang Y. 2009. Uncertainty analysis of
Monin AS, Obukhov AM. 1954. Basic laws of turbulent mixing in the hydrological ensemble forecasts in a distributed model utilising
ground layer of the atmosphere. Trans. Geophys. Inst. Akad. Nauk short-range rainfall prediction. Hydrology and Earth System Sciences
USSR 151: 163–187. 13: 293–303.

Copyright © 2011 John Wiley & Sons, Ltd. Hydrol. Process. 26, 3012–3031 (2012)

Вам также может понравиться