You are on page 1of 21

19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022].

See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
RESEARCH ARTICLE Assessment of Climate Change Impacts on Reservoir
10.1029/2019WR026022
Storage Reliability, Resilience, and Vulnerability Using a
Key Points:
• A multivariate frequency bias
Multivariate Frequency Bias Correction Approach
correction (MFBC) approach is
developed and applied to daily
Ha Nguyen1,2, Rajeshwar Mehrotra1 , and Ashish Sharma1
rainfall and temperature across 222 1
catchments in Australia
Water Research Centre, School of Civil and Environmental Engineering, The University of New South Wales, Sydney,
• The bias‐corrected daily time series New South Wales, Australia, 2Department of Hydrology and Water Resources, Thuyloi University, Hanoi, Vietnam
of rainfall and temperature of six
RCMs for current and future climate
are used to assess the changes in Abstract Raw simulations of global or regional climate models are rarely used in catchment scale
rainfall, temperature, and
hydrological impact assessment or subsequent reservoir storage change assessment studies. Keeping this
streamflows over 222 high‐quality
catchments in Australia in mind, this study uses a frequency bias correction alternative for multiple variables to evaluate the impact
• The impact of climate change on of climate change on reservoir storage reliability, resilience, and vulnerability across Australia. The bias‐
reservoir storage reliability,
corrected time series of daily rainfall and temperature are used as inputs to a hydrological model to derive
resilience, and vulnerability across
Australia is evaluated flows and assess change to reservoir storage attributes. A total of six fifth phase of the Coupled Model
Intercomparison Project climate models dynamically downscaled using the Conformal Cubic Atmospheric
Model are used. Streamflow data of 222 high‐quality catchments in near‐natural conditions across Australia
Correspondence to:
are used and change ascertained. The results for the historical climate show that the multivariate frequency
A. Sharma, bias correction approach outperforms the traditional quantile matching alternative in representing the
a.sharma@unsw.edu.au runoff characteristics related to reservoir storage. For the future climate, the results suggest decrease in the
annual mean runoff for most catchments. The proposed approach leads to a smaller decrease in the standard
Citation: deviation of annual runoff and a reduction in the water supply capability, as indicated by a reduction in
Nguyen, H., Mehrotra, R., & Sharma, A. reliability and resilience and an increase in vulnerability, to meet the demand in comparison to both raw and
(2020). Assessment of climate change
impacts on reservoir storage reliability,
quantile matching‐based climate simulations across for most catchments. Overall, a reduction in water
resilience, and vulnerability using a supply capability to meet a given demand in the future for most regional climate models and catchments
multivariate frequency bias correction is projected.
approach. Water Resources Research, 56,
e2019WR026022. https://doi.org/ Plain Language Summary This article follows on from the considerable work our group has
10.1029/2019WR026022
done to develop a multivariate bias correction alternative that operates in frequency space. In course of
Received 23 JUL 2019
our investigation we had to work on fixing the number of zero rainfall days while dealing with daily rainfall.
Accepted 14 JAN 2020 For historical climate our results are better than the traditionally quantile matching approach in terms of
Accepted article online 15 JAN 2020 reproduction of statistics of relevance in hydrology across all catchments, which shows the utility of the
proposed approach. The multivariate nature of the model makes sure that the bias‐corrected time has a cross
dependence structure that is physically consistent. The application of the approach to estimate the future
changes in the reservoir storage characteristics of the catchments across Australia provides useful
information for use by the Australian water resources managers.

1. Introduction
Water resources assessments of climate change impacts at regional and catchment scales, where most adap-
tation measures are planned and implemented, are of great interest to planners, managers, and the broader
community. A number of studies have attempted to evaluate the possible impacts of climate change on
regional water resources (Demaria et al., 2016; Forzieri et al., 2014; Graham et al., 2007; Mohammed
et al., 2015; Reshmidevi et al., 2017; Sperna Weiland et al., 2010; Sperna Weiland et al., 2012; Wood et al.,
2004). Regional climate models (RCMs) or statistical downscaling models are often used to provide future
projections of the climate variables (e.g., precipitation and temperature) relevant for hydrological assess-
ments (Charles et al., 2004; Stehlı́k & Bárdossy, 2002; Mehrotra et al., 2004; Vrac & Naveau, 2007; Wilby
et al., 1998, 2000; Fowler et al., 2007). It is well known that the raw global climate model (GCM) simulations
contain systematic biases, often making it impractical for use especially in water resources applications.
©2020. American Geophysical Union. Such biases come from varied sources, for example, an imperfect representation of climatic physical pro-
All Rights Reserved. cesses, parameterizations of unresolved subgrid scale processes, and a misrepresentation of land‐

NGUYEN ET AL. 1 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

atmosphere or ocean‐atmosphere feedbacks (Eden et al., 2012; Nahar et al., 2017; Sippel et al., 2016; Troin
et al., 2015), but their systematic nature allows corrections to be effective. Bias correction, therefore, forms
a necessary postprocessing step in almost all climate change impact assessment studies especially for hydro-
logic systems.
It may be noted that the traditional bias correction alternatives can only statistically match select statistical
attributes of raw climate model simulations to observations and cannot correct for the misrepresentations of
physical processes in climate models (Chen et al., 2015; Ehret et al., 2012; Maraun, 2016; Maraun et al., 2017;
Nahar et al., 2017). Inspite of these limitations, many climate change impact assessment studies have shown
the usefulness of bias‐corrected climate data from GCMs or RCMs (or even bias‐correcting GCM‐derived
lateral boundary conditions for RCMs [Rocheta et al., 2017]) to represent the present climate and
subsequently providing more realistic hydrological or other impact model (e.g., agricultural) simulations
(Dosio & Paruolo, 2011; Hagemann et al., 2011; Hagemann et al., 2013; Johnson & Sharma, 2011;
Johnson & Sharma, 2012; Johnson & Sharma, 2015; Piani et al., 2010; Teutschbein & Seibert, 2012;
Watanabe et al., 2012; Wood et al., 2004) for the future.
Bias correction algorithms are often designed to correct specific characteristics of climate variables, vary-
ing from simple correction for means and variance (Ghosh & Mujumdar, 2008; Hay et al., 2000;
Lenderink et al., 2007; Wilby et al., 2004) to more complex approaches designed to correct for quantiles
(Boé et al., 2007; Piani et al., 2010; Wood et al., 2004), or high‐ and low‐frequency variability and
persistence attributes (Haerter et al., 2011; Johnson & Sharma, 2011). The quantile mapping (QM) bias
correction approach is very popular and has been used in many climate change impact studies as it is
relatively straightforward to implement and corrects for the entire distribution of climate time series.
In addition, for daily rainfall, it also corrects for the frequency of dry or wet days and possibly improves
the representation of dry‐ and wet‐spell durations in daily rainfall time series (Piani et al., 2010; Rajczak
et al., 2016; Vrac et al., 2016). However, in water storage‐related applications where low‐frequency varia-
bility and persistence attributes assume importance, QM shows limited advantage as it corrects for biases
at a given time scale and does not specifically correct for the biases at higher aggregated time scales
(Haerter et al., 2011; Mehrotra & Sharma, 2016). One possible solution is to apply a nesting of
corrections of selected statistics at multiple predefined time scales (Haerter et al., 2011; Johnson &
Sharma, 2012; Mehrotra & Sharma, 2016). The nested bias correction is considered as highly efficient
as it corrects for the biases in both distribution and persistence characteristics at multiple selected time
scales. However, biases at other time scales, not included in the bias correction procedure, may not be
corrected. To avoid selecting such specific time scales, (Nguyen et al., 2016; Pierce et al., 2015) suggested
applying bias correction in the frequency domain (FBC), thereby making correction independent of time.
The analysis of the time series in the frequency space makes it more suitable for studying the variability
across different time scales without the need of specifying the time scale (Bloomfield, 2004). However,
due to the intermittent characteristics of daily precipitation (e.g., many days of no rain) application of
FBC to daily precipitation is not straightforward (Nguyen et al., 2016, Pierce et al., 2015). Recently,
Nguyen (2018) proposed a modified frequency‐based bias correction framework for correcting the biases
in daily precipitation and named it as FBCd. The FBCd approach combines the advantage of the FBC to
account for the low‐frequency variability in the bias‐corrected time series and the strength of QM to
account for the high‐frequency variability and dry‐ or wet‐day frequency. FBCd was found to correct
effectively both high and low‐frequency variability in daily precipitation. In hydrology and water
resources‐related applications, for example, reservoir operation and drought management, low‐frequency
variability plays an important role and use of FBCd provides better results (Nguyen, 2018).
A catchment‐scale hydrological model usually requires multiple climate variables as input (such as daily
precipitation and temperature/evaporation). Use of a univariate bias correction approach such as FBCd to
correct for biases in time series of individual variables may lead to suboptimal performance of the hydrolo-
gical model. Nguyen et al. (2018) proposed a multivariate extension of FBC. Following Nguyen et al. (2018),
we present here a multivariate extension of FBCd, called MFBCd, which, in addition to temporal depen-
dence, also considers the dependence across multiple atmospheric variables. In summary, MFBCd is aimed
at correcting across variables dependence biases, temporal dependence biases at multiple time scales and the
high‐frequency variability and dry‐ or wet‐day frequency biases.

NGUYEN ET AL. 2 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

Figure 1. Methodology flowchart.

Our results include an assessment of the performances of the multivariate frequency bias correction, MFBCd,
traditional QM and raw RCM data. The study examines the outcome of these data sets on water security of
potential water supply systems across Australia. Reservoir storage attributes of 222 catchments over
Australia are used for this purpose. In the first part of the study, we compare raw RCM and bias‐corrected
simulations of precipitation, temperature, streamflow, and reservoir storage attributes for present climate
conditions. In the second part, we assess climate change impacts on reservoir storages by evaluating the
future projections of changes in streamflow and reservoir storages driven by raw and bias‐corrected data sets.
The rest of the paper is organised as follows. Section 2 presents the details of the methodology and data used.
Section 3 discusses the performances of MFBCd and QM in improving the representations of runoff and
reservoir storage characteristics for the historical climate. Section 4 compares the changes in the runoff
characteristics and reservoir storage attributes of raw and bias‐corrected simulations for the future climate.
Finally, Section 5 presents discussions and conclusions drawn from this study.

2. Methodology, Study Catchments, and Data Sets


The methodology adopted in this study is summarized in Figure 1. The RCM‐simulated raw daily precipita-
tion and temperature time series are bias corrected using the MFBCd and QM approaches. A hydrological
model calibrated using observed data is forced with observed, raw (or uncorrected), and bias‐corrected daily
time series of precipitation and temperature to simulate daily streamflow time series. The observed and
simulated streamflows are fed to a reservoir with a defined water demand to compute reservoir
storage characteristics.
An important aspect in reservoir systems planning and performance is to assess the system performance
under wide range of possible demands and hydrologic conditions that are expected to occur during the oper-
ating life of the reservoir. However, this information is not available for catchments where no dams exist. In
the absence of such information, an alternative could be to go with some reasonable assumptions about
reservoirs demands and storages and operate reservoirs using simple relationships. In order to be consistent
in our analysis across all catchments and considering the fact that we don't have reservoirs in most of the
catchments considered, we decide to go ahead with hypothetical reservoirs. It is common to fabricate these
reservoirs by assuming a reasonable demand and keeping it as a function of draft and reliability (Ehsani
et al., 2017). Adoption of demand equal to 50–90% of draft and storage necessary to meet demand at 20–
90% of reliability is common. Considering the fact that our aim is to assess changes across the entire country,
to allow a fair assessment of reservoir storage performance across varying catchments in varying climate
zones, the size of hypothetical storage is set as 90% reliability of meeting varying demands (drafts equal to
75% and 85% of the annual mean streamflow of the catchment). Also, in order to keep the analysis

NGUYEN ET AL. 3 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

Figure 2. Hydrologic Reference Stations (HRSs): (a) Locations of 222 HRSs, (b) distribution of catchment areas, (c) cumu-
lative distribution of Nash‐Sutcliffe efficiency (NSE) values of calibrated GR4J hydrological model across 222 HRSs.

simple, we have not included cascading reservoir effect, if any, in our analysis, mostly because we do not
have flow observations at multiple points in the catchment.
Characteristics of simulated runoff and reservoir storages attributes are analyzed to determine the implica-
tions of MFBCd and QM in modeling the streamflow and water availability for historical and future climates.
Details of study catchments, data sets, and performance metrics used in this study are described in the
following sections.

2.1. Data Sets


2.1.1. Study Catchments
We use streamflows for 222 high‐quality near‐natural catchments across Australia for our study. These
catchments represent Australia's network of Hydrologic Reference Stations (HRSs) developed by the
Bureau of Meteorology and are available at http://www.bom.gov.au/water/hrs/. These 222 catchments
cover all of Australia (Figure 2a), with areas ranging from 4.5 to 232,846 km2 (Figure 2b), covering different
climatic conditions from the temperate regions of the south to the arid expanses of the interior and to the
tropical regions of the north. The associated results presented in the Figure 2c are discussed in more

NGUYEN ET AL. 4 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

detail in the next section. For a more detailed description of these 222 HRSs, readers are referred to Zhang
et al. (2016) and Ajami et al. (2017).
2.1.2. Climate Data Sets
Observed and RCM‐simulated climate data sets used in this study consist of daily precipitation and tempera-
ture data. Observed gridded daily precipitation and temperature time series are obtained from the Australian
Water Availability Project (AWAP) prepared by the Australian Bureau of Meteorology (Raupach et al., 2009,
2012). The AWAP data sets cover all of Australia at a resolution of 0.05° × 0.05° and are available at http://
www.bom.gov.au/climate/maps/#tabs=About‐maps‐and‐data. These data sets are used to calculate
catchment averaged rainfall and temperature time series over 222 catchments.
RCM data sets using the Conformal Cubic Atmospheric Model (CCAM) forced with six Global Climate
Models (GCMs) (namely, ACCESS1‐0, CCSM4, CNRM‐CM5, GFDL‐CM3, MPI‐ESM‐LR, NorESM1‐M) over
the AustralAsia domain covering Australia and surrounds are available at a resolution of ~50 km (see http://
cordex‐australasia.wikidot.com/experiments) (Evans, 2011). These data sets are generated as a part of the
Coordinated Regional Downscaling Experiment (CORDEX) program of the World Climate Research
Program and are available at https://www.nsc.liu.se/storage/esgf‐datanode/ (Giorgi et al., 2009).
For the assessment of historical climate simulations, daily time series of climate variables from 1971 to 2014
is extracted for both observed and RCM‐simulated data sets. In fact, RCM‐simulated data sets for the
historical climate period are only available from 1971 to 2005; however, in order to have a longer time series
and to match its length with the observations (for the purpose of cross‐validating the bias correction
approaches), we concatenate this RCM‐simulated historical period data with data from 2006 to 2014
obtained from Representative Concentration Pathway (RCP) 8.5 scenario simulations. For the assessment
of future simulations, the RCP 8.5 scenario simulations for the period from 2056 to 2090 are adopted.

2.2. Bias Correction of RCM Simulations


The following summarizes the implementation procedure of QM and MFBCd bias correction approaches.
In QM, the entire distribution of modeled time series is corrected to match that of observed time series, using
either parametric or nonparametric approaches. The QM approach used here is based on a nonparametric
approach, also known as empirical QM, (EQM) following the formulation of Themeßl et al. (2012):
n −1
   −1
  o
X qm;val
d ¼ X raw;val
d þ F obs;cal F raw;cal X raw;val
d −F raw;cal F raw;cal X raw;val
d (1)

where X represents a variable (e.g., precipitation or temperature) and F(.) indicates the cumulative
distribution function with F−1(.) its inverse. The superscripts obs, raw, and QM indicate the observed,
raw, and QM‐corrected values. The superscripts cal and val indicate the calibration and validation (or future)
time periods. The subscript d indicates daily time scale.
We apply QM to model daily precipitation and temperature data separately for each of the 12 calendar
months. The empirical cumulative distribution functions for each month are constructed using pooled daily
data falling within a sliding 3‐month window centered on the month of interest. In addition, an extrapola-
tion of the correction function following Themeßl et al. (2012) is applied for values outside the range of
calibrated values.
In order to effectively represent dry days in daily precipitation, zero values in the observed and modeled time
series are replaced with uniform random numbers between 0 and a small threshold (e.g., 0.01 mm) before
applying QM. After bias correction, values below the threshold are set equal to zero (Cannon, 2017;
Cannon et al., 2015; Vrac et al., 2016). This procedure corrects for biases in wet or dry frequencies even
though the model underestimates or overestimates numbers of wet days in comparison to observations.
MFBCd is based on the concept of a multivariate frequency‐based bias correction (MFBC). Use of the fre-
quency domain provides a simple mean of correcting low‐ and high‐frequency variability biases and results
in an improved representation of observed variability and persistence in the corrected time series over a wide
range of time scales. The MFBC approach matches the univariate spectrum of individual time series to that
of observed time series and then matches phase differences of pairs of modeled time series to those of pairs of

NGUYEN ET AL. 5 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

observed time series. Nguyen et al., 2018 describe the procedure in detail. The general algorithm of MFBCd
includes the following steps:
a Apply QM to correct for entire distribution of daily precipitation and temperature for modeled data.
Aggregate/average the QM bias‐corrected time series to form monthly time series of these variables.
b Aggregate/average the raw RCM time series of precipitation and temperature to form raw monthly time
series of these variables. Apply MFBC on the raw monthly time series.
c Adjust MFBC corrected monthly time series as obtained from step (b) to match the QM‐corrected and
aggregated time series (obtained from step [a]). Number of wet days in the precipitation time series is
adjusted using stochastic precipitation occurrence adjustment algorithm (Nguyen et al. (2019)). The pro-
cedure adds or removes a subset of wet days in the precipitation sequence.
d Form final bias‐corrected daily precipitation and temperature time series using the following scaling
equation:

XQM
XMFBC
d
d
¼ XdQM XMFBC
m ;(2),
m

where X represents a variable (e.g., precipitation or temperature). The superscripts QM, MFBC, and MFBCd
indicate the QM‐, MFBC‐, and MFBCd‐corrected values. The subscripts d and m indicate daily and monthly
time scales. Note that, for precipitation, XQM d and Xqm
m are adjusted following stochastic
precipitation occurrence adjustment as mentioned above.

2.3. Runoff Simulations


A lumped conceptual daily rainfall‐runoff model—GR4J (Perrin et al., 2003)—is used to simulate daily
streamflow forced with different input options of climate data (observed, raw, and bias‐corrected climate
time series for historical and future periods).
The daily GR4J model is a simple and popular model and has been used successfully in many catchments
over the world (Coron et al., 2017; Pagano et al., 2010; Perrin et al., 2003; Teng et al., 2015). The model
requires daily time series of precipitation and potential evapotranspiration as an input and only four para-
meters to define its structure. Following Hamon (1961), daily potential evapotranspiration (PET) time series
is estimated using daily mean air temperature and radiation. Radiation is calculated based on daytime length
(as a function of latitude and declination) (Lu et al., 2005). In a study, Oudin et al. (2005) showed that the
PET estimated using temperature and radiation provided the best streamflow simulations. The GR4J model
is calibrated separately for each of the 222 catchments using observed daily streamflow data and catchment‐
averaged daily precipitation and PET data. The parameters of GR4J model for each catchment are optimized
using the shuffled complex evolution global optimization algorithm (Duan et al., 1993) based on the maxi-
mization of the Nash‐Sutcliffe efficiency (NSE) as an objective function. The calibration results show about
90% of the catchments with NSE values greater than 0.6 and about 77% with NSE values greater than 0.7
(Figure 2c). Following these results, subsequent analysis only includes the catchments with NSE greater
than 0.7 (resulting into about 77% of the HRSs).
To assess the impact of the two bias correction approaches on runoff and ensuing implications on reservoir
storage, we force the calibrated GR4J model with observed, raw, and bias‐corrected data to obtain different
simulated runoff time series. These are termed as
i Qsim_obs—obtained using the observed daily precipitation and temperature time series,
ii Qsim_raw—obtained using the raw (or uncorrected) RCM‐simulated daily precipitation and tempera-
ture time series,
iii Qsim_qm—obtained using the QM‐bias‐corrected daily precipitation and temperature time series,
iv Qsim_mfbcd—obtained using the MFBCd‐bias‐corrected daily precipitation and temperature time series.

2.4. Precipitation, Temperature, and Runoff Characteristics Considered


Given the objective of this study, a number of key climate and runoff attributes related to reservoir storage
are examined. The key attributes for the climate (precipitation and temperature) data include mean and
standard deviations, persistence attributes (measured by lag‐1 autocorrelation), and cross‐dependence

NGUYEN ET AL. 6 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

attributes (measured by lag‐0 cross‐correlations) between precipitation and temperature at multiple time
scales (daily, monthly, and annual).
For runoff, we focus on the mean, standard deviations, and the high and low flows (defined as the flow
exceedance at 5th and 95th percentiles extracted from the flow duration curve) at multiple time scales (daily,
monthly, and annual).

2.5. Reservoir Storage Performance Measures


We reiterate that one of the key aims here is to analyze climate change impacts on reservoir storages of con-
siderable relevance given the general scarcity of water in arid countries such as Australia (Zarch et al., 2015).
As the reservoirs are hypothatical and indicative of likely changes in other storages that exist or are planned,
we adopt demands equal to a 75% and 85% of draft (expressed as a percentage of annual mean inflow) and
decide reservoir storage as the storage necessary to meet this demand at 90% reliability. Then, we adopt a
commonly used heuristic operating policy, so‐called standard operating policy, to compute the reservoir
storage performance characteristics (McMahon et al., 2007; Turner & Galelli, 2016). The characteristics of
reservoir storage performance are defined by three performance metrics following Hashimoto et al. (1982)
and McMahon et al. (2007) as described below. This analysis is performed on monthly time series.
The first metric is the reliability of monthly time series, which is defined as the probability (proportion of
months during the simulation period) that a reservoir can meet the target demand within the
simulation period.

Ns
Rt ¼ (3)
N

where Rt is the monthly reliability, Ns is the number of months during which the demand was met, and N is
the total number of months in the simulation.
The second metric is resilience, which measures the ability of the reservoir to recover after a failure.

fs
φ¼ (4)
fd

where φ is resilience, fs is the number of continuous sequences of failure periods, and fd is the total duration
of the failures.
The last metric is vulnerability (also called dimensionless vulnerability) and measures the average of the
maximum shortfalls occurring in each of the continuous failure periods.

fs  
∑ max sj
η ¼ k¼1 ; (5)
Df ×f s

where η is vulnerability, sj is the volumetric shortfall during the jth continuous failure sequences, and Df is
the target draft during failures.

3. Model Performance Evaluation Using Present Climate Simulations


In this section, we compare and evaluate the performance of both bias correction approaches in capturing
important characteristics of precipitation, temperature, and runoff. In addition, we also look at the perfor-
mance characteristics of the streamflow to meet a given demand and reliability (75% and 85% drafts and
90% reliability). The performance evaluation is based on cross‐validation by dividing the historical period
(44 years) into four blocks of 11 years, and bias correcting the climate time series of each block at a time with
the model being formulated using the remaining data of three blocks (Mehrotra & Sharma, 2006). Hence, all
results presented here are reflective of what possibly can be obtained in a different setting (for example, in
the future), as the assessment is based on a period not used in the model specification.

NGUYEN ET AL. 7 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

We first look at the model performance at individual catchments. Due to the large number of results and
the fact that they convey similar information, we choose to present the detailed results for two RCMs,
one driven by ACCESS1_0 (denoted as CCAM‐ACCESS) and the other driven by CCSM4 (denoted
as CCAM‐CCSM4).
Figures 3 and 4 show the scatterplots of observed and modeled statistics of the raw and bias‐corrected rainfall
and temperature data for the two selected RCMs across all the catchments. A dot on the plot indicates a
catchment. On these plots, for each statistic, we also show mean absolute error (MAE) between the statistics
of raw and bias‐corrected data with that of observed data averaged over the catchments. These plots show
that both QM and MFBCd generally improve the representation of statistical attributes of rainfall and tem-
perature over the raw RCM simulations.
While both QM and MFBCd perform similarly in term of reducing the biases in the day‐to‐day and month‐to‐
month variability, their performance in reproducing the observed variability at annual time scale (interann-
ual standard deviation) is different. MFBCd consistently leads to substantial reductions of the biases in the
annual standard deviation of both rainfall and temperature, while QM offers very limited improvement in
the statistic for the both RCMs, except for the annual standard deviation of rainfall for the CCAM‐CCSM4
model. Surprisingly, the QM approach leads to marginally smaller MAE of the annual standard deviation
rainfall for the CCAM‐CCSM4 model than the MFBCd approach; however, for the catchments with the lar-
ger bias in this statistic, the MFBCd approach still shows larger improvements than the QM as shown in
Figure 4c. Further analysis of the structure of the biases in the rainfall variability reveals that for both
RCMs the biases in the daily and monthly time scales are quite similar, resulting in similar performances
for both bias correction approaches. However, the biases in the annual variability are similar for CCAM‐
CCSM4 but very different for CCAM‐ACCESS for most of the catchments. This explains why the QM
approach leads to improvement for the annual standard deviation of rainfall for CCAM‐CCSM4 and not
for the CCAM‐ACCESS model, while the MFBCd approach provides consistent performance for both
RCMs. This implies that if the biases in the variability at different time scales are quite alike, one simple bias
correction, like QM, applied at a single time scale (e.g., daily) may reproduce well both short‐term and long‐
term variabilities as a more complex or nested bias correction model is not needed.
Turning to the auto‐ and cross‐dependence statistics of rainfall and temperature, the MFBCd approach gen-
erally leads to improved estimates of lag‐1 autocorrelations and lag‐0 cross‐correlations of simulated preci-
pitation and temperature (smallest MAE values) at monthly and annual time scales in comparison to QM.
This is to be expected as the structure of MFBCd is built to correct for biases in the auto‐ and cross‐spectra
in the frequency domain, and thus leads to improvement in their representations in the time domain.
However, we note that if the raw model shows minor biases in these statistics, the two bias correction
approaches may lead to similar results in their performance, as the case of lag‐1 autocorrelation of monthly
temperature and lag‐0 cross‐correlation between monthly rainfall and temperature time series.
We now examine the performance of the two bias correction approaches on runoff characteristics. Similar to
Figures 3 and 4, Figures 5 and 6 show the statistics of raw and bias‐corrected streamflows (Qsim_raw,
Qsim_qm, and Qsim_mfbcd) for CCAM‐ACCESS and CCAM‐CCSM4 models plotted against the correspond-
ing observed statistics (Qsim_obs).
Please note that in all the results related to storage characteristics reported hereafter, runoff time series
simulated by the calibrated GR4J model using observed precipitation and temperature data is considered
as representative of observed flow series to avoid attribution due to biases in the hydrological model used.
First, we look at the streamflow variability, represented here in the form of standard deviation. Both bias
correction approaches reduce the biases in the standard deviation of streamflow at different time scales.
The impact of these approaches in reducing variability biases generally follows their performance in cor-
recting the bias in the rainfall and temperature standard deviation. For example, the MFBCd approach
greatly reduces the bias in the interannual standard deviation for both RCMs resulting into the best esti-
mates of this statistic. However, the QM approach reduces this bias for the CCAM‐CCSM4 model only,
similar to what we observed for annual standard deviation of rainfall. For the statistics related to high‐
and low‐flow characteristics (the 5th and 95th percentile flow values) at different time scales, the MFBCd
approach generally leads to a better performance over QM approach, especially at the annual time scale
for both RCMs (Figures 5 and 6).

NGUYEN ET AL. 8 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

Figure 3. Scatterplots of statistics of observed, raw, QM, and MFBCd bias‐corrected rainfall and temperature time series for the RCM driven by ACCESS1_0. The
simulated statistics in these plots are shown on the vertical axis, while the observed ones are shown on the horizontal axis. Circles on these plots represent
catchments with different colors showing results for raw (gray), QM (pink), and MFBCd (green). The MAE value of each statistic is shown on each plot at the top left.

NGUYEN ET AL. 9 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

Figure 4. As in Figure 3 but showing results for the RCM driven by CCSM4.

Finally, we examine the performance of both bias correction approaches in simulating the reservoir storage
performance characteristics. Figures 7 and 8 provide box plots of three reservoir storage performance metrics
for the two selected RCMs, demands, and all catchments. Box plots show the distribution of these character-
istics across catchments. Box plots are quite similar for both demands (first column vs. second column of

NGUYEN ET AL. 10 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

Figure 5. Scatterplots of runoff characteristics simulated by the calibrated GR4J model using the observations with those using raw, QM, and MFBCd bias‐corrected
data for the CCMA‐ACCESS model. The simulated statistics on these plots are shown on the vertical axis, while the observed ones are shown on the horizontal axis.
Values are plotted on a log‐log scale (with base of 10) for the purpose of visualizing results across catchments. Circles on these plots represent catchments with
different colors showing results for raw (gray), QM (pink), and MFBCd (green). The MAE value of each statistic is shown on each plot at the top left.

Figures 7 and 8). As expected, QM and MFBCd approaches reproduce these performance metrics better than
the raw RCM results for most of the catchments. Irrespective of the demand, the QM approach often leads to
higher estimates of reliability and resilience and lower estimates of vulnerability compared to the MFBCd
approach. This implies that the QM approach provides lower likelihood, lower severity of shortfall, and
quicker recovery of reservoirs from failure than the MFBCd approach. That would be beneficial if the QM
approach consistently reproduces these performance measures better than the MFBCd approach.
However, this is not always the case. For the CCAM‐ACCESS model, the MFBCd approach mostly leads
to better estimates of these reservoir‐storage performance measures (with lowest MAE values on average).
Specifically, MFBCd provides reliability values closer to the target (90%) and close to zero biases
(simulated − observed) in the resilience and vulnerability for most catchments (Figure 7). For the CCAM‐
CCSM4 model, although QM overestimates the design reliability for most catchments, on average it
provides the lowest MAE for reliability for both demands. As mentioned earlier, the QM approach
corrects quite well the low‐frequency variability bias in the rainfall and runoff time series for these
catchments due to the specific structure of the biases in the rainfall time series for this RCM. The MFBCd
approach overcorrects the low‐frequency variability bias in the rainfall time series, resulting into the
underestimation of the reliability for these catchments. However, when the shortage actually assumes
significant proportions, MFBCd provides estimates of the reservoir storage flexibility in recovering from a

NGUYEN ET AL. 11 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

Figure 6. As in Figure 5 but showing results for the RCM driven by CCSM4.

failure (measured by resilience) and the severity of failure (measured by vulnerability) closest to
observations, while QM leads to higher resilience and lower vulnerability for most catchments (Figures 8b
and 8c). This suggests the advantage of the MFBCd approach is in correcting for low‐frequency variability
in the climate input time series over the QM approach and hence providing a better representation of
reservoir storage performance.
The performance results presented so far are for two selected RCMs with the biases exhibiting markedly
different structures. To assess bias correction models’ performance across all RCMs and catchments,
Figure 9 shows the averaged model performance of the two bias correction approaches in capturing
the important characteristics of rainfall, temperature, and runoff. The average model performance is
quantified by MAE between the attributes of raw and bias‐corrected data with those of observed data
averaged over all RCMs and catchments. In order to present the MAE values of different attributes on
the same plots, we divide the MAE values derived from the bias‐corrected data by those derived from
the raw RCM data, so‐called relative MAE. For simplicity of presentation, we do not present the results
of statistics where raw RCM simulations show small biases, resulting in similar performance between
the two bias correction approaches. As anticipated, with the exception of monthly standard deviation,
MFBCd performs well across all statistics considered. In general, the results of Figure 9 show that the
MFBCd approach performs better than the QM approach, a difference which is especially amplified for
characteristics related to low‐frequency variability. This is highly relevant to water resources manage-
ment as variability is a key driver in such cases.

NGUYEN ET AL. 12 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

Figure 7. Box plots of reservoir storage performance metrics derived from the observed, raw, QM, and MFBCd bias‐corrected data for the RCM driven by
ACCESS1_0 for 75% (left) and 85% (right) drafts. In top row (7a), we show reliability estimates, while in middle and bottom rows (7b and 7c), we show biases
(raw or bias‐corrected values − observed values) in resilience and vulnerability estimates, respectively. The MAE value of each performance metric is shown along
the title of each plot.

4. Implications for Future


The future changes in the raw and bias‐corrected RCMs results are computed with respect to the raw and
bias‐corrected historical climate results, respectively. For the sake of simplicity, only annual mean and
annual standard deviation are used for comparisons of future changes in the rainfall, temperature, and run-
off characteristics obtained from the three cases (the raw RCM simulations and the two bias correction
approaches). In addition, future changes in the runoff and storage characteristics are evaluated in terms
of the changes in the three reservoir storage performance measures, namely, reliability, resilience, and vul-
nerability. Figure 10 shows the percentage changes in annual means and annual standard deviations of rain-
fall, temperature, and runoff time series obtained for the three cases across all catchments. The plots in the
first three columns show changes (an increase in blue circle, decrease in red circle, and no change in gray
circle) and relative magnitude of these changes (larger circles imply higher magnitudes) at individual catch-
ments averaged over the six RCMs (RCM average) for the raw RCM simulation (the first column), QM (the
second column), and MFBCd (the third column) bias‐corrected simulations. The plots in the last column
summarize the spatial variation of the percentage changes in the statistics of RCMs average and individual
RCMs. As observed in Figure 10, although the three cases (Raw, QM, and MFBCd) produce similar spatial
patterns of the changes for the annual mean rainfall, temperature, and runoff, there are some differences
in the results for the annual standard deviation statistic across the three cases.
For the annual mean rainfall, the RCMs average results for all the three cases suggest a decrease in this sta-
tistic for most of the catchments, namely, for the catchments located in Tasmania, South West coast, North
East coast, and the south of southeast Australia; an increase in this statistic is only expected for the

NGUYEN ET AL. 13 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

Figure 8. As for Figure 7 but showing results for the RCM driven by CCSM4.

catchments in the north of southeast Australia and a few catchments in north Australia (Figures 10a–10d).
However, there are significant differences across RCMs results despite the use of Raw, QM, or MFBCd
(Figure 10d). More specifically, the three RCMs (driven by ACCESS1_0, GFDL, and MPI) indicate a decrease
in the annual rainfall, while other two RCMs (driven by CCSM4 and CNRM) indicate an increase in this
statistic for almost all catchments, and the last RCM (driven with NorESM1) provides results similar to those
for the RCM‐average.
For the annual standard deviation of rainfall, the spatial pattern of the changes derived for the raw RCM
simulations is closer to that obtained for the MFBCd approach than for the QM approach (Figures 10e–
10g). For example, both the raw and MFBCd‐corrected simulations lead to an increase in this statistic for
the catchments in southeast Australia and most catchments in Tasmania, but the QM‐corrected simulations
project a decrease for these catchments. In addition, the median changes suggest a small increase in this
statistic for both the raw and MFBCd‐corrected simulations but a decrease for the QM corrected
simulations (Figure 10h).
For temperature, the annual mean and annual standard deviation are expected to increase in the future
period for all the catchments under all the three cases (Figures 10i–10o). We again note that the QM‐
corrected simulations lead to bigger differences in the projections of the change in the annual standard
deviation across the catchments from different RCMs compared to the two other cases.
The changed patterns of runoff generally follow the projected changes in rainfall and temperature. For
example, with decrease in annual mean rainfall and increase in annual mean temperature for most of the
catchments, we expect a decrease in the annual mean runoff for all the three cases. In addition, the three
RCMs driven with ACCESS1_0, GFDL, and MPI more or less agree with each other about the range and
median of the projected changes for the annual mean runoff, while the two RCMs driven with CCSM4

NGUYEN ET AL. 14 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

Figure 9. Comparison of overall performance of two bias correction approaches in reproducing the attributes of rainfall,
temperature, and runoff (in cross‐validation) across all the RCMs and catchments considered for 75% and 85% drafts.
Relative mean absolute error (RMAE) values are obtained by dividing the MAE values of the bias‐corrected data by the
raw RCM data; mstd and ystd denote monthly and annual standard deviations, respectively; mlag1 and ylag1 denote
monthly and annual autocorrelations, respectively; yr0 denote annual cross‐correlation; yQ5 and yQ95, the flow excee-
dance at 5th percentile and 95th percentile respectively, extracted from the respective daily, monthly, and annual flow
duration curves.

and CNRM provide these projection results more closely to each other across the catchments in spite of the
use of either bias correction approaches.
While the three cases generally produce the similar median changes and range in the annual mean runoff
across the catchments, there are differences in the projected changes of the standard deviation of annual
runoff among them. For example, the median of the RCMs average change suggests that the MFBCd
approach leads to the smallest decrease in the annual standard deviation, followed by the raw RCM simula-
tions, and then the QM approach. The difference in the projections is found mostly in southeast Australia
where the MFBCd approach leads to an increase in the annual standard deviation for many catchments, par-
ticularly in the south of southeast Australia, while the QM approach leads to a larger decrease (larger circles
in Figure 10v) in this statistic for most catchments across Australia.
We now examine whether the three cases (Raw, QM, and MFBCd) behave differently for the water supply
capability of a regulated hypothetical reservoir in the future. Figure 11 shows the absolute changes in the
reservoir storage performance characteristics (reliability, resilience, and vulnerability) calculated for these
cases. Similar to Figure 10, we present here the results of individual catchments averaged over six RCMs
(the plots in the first three columns of Figure 11) and the spatial distribution of the changes in the RCMs
average and individual RCMs (the plots in the last column of Figure 11). The spatial patterns of future
changes for the reliability and resilience across the catchments for all RCMs are like those for annual mean
runoff. For example, decreases in the reliability and resilience are projected for almost all the catchments,
and these patterns are like those of annual mean runoff. One may expect increased reliability and resilience
associated with the projected decrease in the standard deviation of annual runoff. However, as annual runoff
is projected to decrease in the future, the decrease of the standard deviation of annual runoff does not offset
this decrease. Comparing the results of these two performance metrics acorss different RCMs, the three
RCMs (driven with ACCESS1_0, GFDL, and MPI) project the larger decreases than the other three RCMs

NGUYEN ET AL. 15 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

Figure 10. The future projections of percentage changes (the difference of future and historical values divided by the historical values) in annual means and stan-
dard deviations of rainfall, temperature, and runoff obtained from Raw, QM, and MFBCd across all the catchments. The percentage changes averaged for all the six
RCMs (RCMs average) at individual catchments are presented in the first three columns. Circles on these plots represent catchments with different colors showing
an increase (blue), decrease (red), or no change (gray) in the characteristics. In addition, the size of circles represents the relative magnitude of the percentage
changes (a bigger circle means a larger change). The spatial distributions of percentage changes for each RCM are presented by the boxplots in the last column.

NGUYEN ET AL. 16 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License

Figure 11. As for Figure 10 but showing results for the future projections of absolute changes (future values − historical values) for reservoir storage performance

17 of 21
10.1029/2019WR026022
Water Resources Research

metrics for 75% (top) and 85% (bottom) drafts.

NGUYEN ET AL.
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

(driven with CCSM4, CNRM, and NorESM1) following the similar pattern of the projected results of annual
mean runoff. The decrease in the reliability and resilience for all the three cases imply the increased
frequency of system failure and decreased flexibility of recovering from a failure. Comparing the results
across the three cases, we note that the MFBCd approach leads to larger decreases in the reliability, smaller
decreases in resilience, and smaller increases in vulnerability compared to the raw and QM cases for most
RCMs and catchments (larger circles in Figures 11c and 11g). This is likely because the MFBCd approach
reproduces the observed variability to a markedly better extent than QM, as was demonstrated in the
cross‐validation results presented earlier in section 3.
Finally, for the vulnerability estimates, the RCMs average results for all three cases suggest an increase in the
vulnerability for the most catchments. Although there are differences in the projected results across the six
RCMs, the median and the range of these absolute changes suggest an increase in the vulnerability in the
future. We observe that the MFBCd approach for the RCMs average and for most RCMs leads to higher
increases in the vulnerability compared to the remaining two cases. This behavior is similar to the reliability
and resilience results mentioned above.

5. Discussion and Conclusions


This study investigates the implications of QM and MFBCd bias correction approaches on reservoir storages
for historical and future climate. The assessment is based on the results of CCAM RCM forced with six GCMs
for 222 high‐quality streamflow stations in near‐natural catchments across Australia.
The cross‐validated results for the historical climate show that both QM and MFBCd bias correction
approaches lead to a better representation of streamflow and reservoir storage characteristics over raw
RCM simulations. The MFBCd approach often leads to improved estimates of streamflow low‐frequency
variability and provides a better reproduction of the reservoir storage characteristics than the QM approach.
The QM approach is often found to provide increased reliability and resilience and decreased vulnerability
than the corresponding observed estimates.
For the future climate, both raw and bias‐corrected simulations of RCMs suggest reductions in the means
and standard deviations of annual streamflows over majority of catchments. The results also indicate
increased vulnerability and decreased reliability and resilience in meeting a targeted demand for all the three
cases. For majority of RCMs and catchments, while the patterns of the projected changes for the annual
mean show no noticeable differences across the three cases (raw, QM, and MFBCd), we find differences in
the projected changes for the standard deviation of annual streamflows across the three cases. MFBCd
approach generally leads to higher estimates of annual standard deviation, in other words, smaller decreases
in this statistic in the future, compared to the Raw and QM cases. As a result, in comparison to QM, MFBCd
approach results in a larger reduction of water supply capability meeting a given historical demand in the
future for most RCMs and catchments.
We note that the results presented here are not free from the limitations associated with the most bias
correction approaches discussed in the literature (Maraun, 2016; White & Toumi, 2013; Ehret et al., 2012).
Assessment of any bias correction approach for future climate is difficult as there is no data for
verification and is even harder to justify the fundamental assumption that the bias correction procedures
are stationary. This is the main weak point of any bias correction approach and emphasizes the need to
improve RCMs (and GCMs) in order to reduce or eliminate the biases altogether.
Despite these limitations, bias correction remains popular and an essential post processing tool of any
climate change impact assessment study (e.g., Bennett et al., 2014; Chen et al., 2013; Christensen et al.,
2008; Hagemann et al., 2013; Macadam et al., 2016). This study highlights the importance of bias correction
of RCM‐simulated precipitation and temperature time series with a special focus on the streamflow
simulations and reservoir storage assessment.
This study also highlights the importance of across‐time scale variability biases in climate simulations used
for water resources system assessments. We show that if the biases in across‐time scale variability are not
aligned, a simple bias correction, such as QM, applied at a single pre‐chosen time scale (e.g., daily) cannot
capture the biases at other time scales (e.g., annual). The differences between the two bias correction
approaches are mostly related to their ability to simulate the across‐scale variability in precipitation and

NGUYEN ET AL. 18 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

temperature, which is when translated into streamflow show differing results on the performance
characteristics of engineering storage infrastructures.
These results are further strengthened by the fact that the study utilizes outputs of multiple RCM
simulations. We show that the RCM simulations do not always agree on the magnitude or even the sign
of changes in the streamflow and reservoir storages. In this study, combined multimodel projections of
the future changes are obtained by arithmetically averaging the results across all the RCMs. Further work
would quantify and include the uncertainties associated with different RCMs into hydrological climate
impact studies or adopt a suitable weighted ensemble average scheme for improving multimodel climate
change projections (Reshmidevi et al., 2017; Woldemeskel et al., 2014; Woldemeskel et al., 2016).
Rather than using bias‐corrected meteorologic variables and a rainfall‐runoff model to obtain the flows as
has been adopted in this study, an alternative approach could be to directly bias correct the raw flows derived
from RCMs or again apply bias correction on the flows obtained after simulation. The former approach is
used commonly, possibly because the ensuing model has lesser complexity and also because it is easier to
obtain the meteorologic variables from the RCMs/GCMs, and we have more confidence in these variables
in comparison to flows which forms a secondary output of the models. Moreover, flows are known to be
influenced by time‐varying catchment and climate characteristics (Burt, 2013; Lv et al., 2019), which are
unlikely to be captured well either by RCM/GCM or by any bias correction model and therefore are likely
to have larger uncertainty in a future climate. A possible future research topic would be to compare our
results with those obtained by applying the bias correction again on the simulated flows with a special focus
on the reservoir operation and drought management.
The use of one simple conceptual hydrological model—GR4J model—could also form one of the limitations
of this study as some studies (e.g., Hagemann et al., 2011; Hagemann et al., 2013; Sperna Weiland et al., 2012)
showed that the choice of a hydrological model might have some impacts on the future projections of
streamflow. It might be of interests to repeat the analysis and compare the results using different
Acknowledgments
We acknowledge the World Climate hydrological models.
Research Programme's Working Group
on Regional Climate, and the Working Finally, our assessment of reservoir storage performance is limited to the assumed reservoir capacities to
Group on Coupled Modelling, former meet 75% and 85% drafts with 90% reliability and following the standard operating policy. Different
coordinatingbody of CORDEX and performance characteristics of reservoir storages would be expected if different estimates of reservoir storage
responsible panel for CMIP5. We also
thank the CORDEX‐AustralAsia mod- capacities, changed demands, and altered operation policy are adopted. This would be the subject of
eling groups and their respective insti- future work.
tutions for their roles in producing and
making available their model output. The MATLAB code used in this research can be downloaded from the following website: http://www.
We also acknowledge the Earth System hydrology.unsw.edu.au/download/software.
Grid Federation infrastructure an
international effort led by the U.S.
Department of Energy's Program for
Climate Model Diagnosis and
Intercomparison, the European
References
Network for Earth System Modeling, Ajami, H., Sharma, A., Band, L. E., Evans, J. P., Tuteja, N. K., Amirthanathan, G. E., & Bari, M. A. (2017). On the non‐stationarity of
and other partners in the Global hydrological response in anthropogenically unaffected catchments: An Australian perspective. Hydrol. Earth Syst. Sci., 21, 281–294.
Organisation for Earth System Science Bennett, J. C., Grose, M. R., Corney, S. P., White, C. J., Holz, G. K., Katzfey, J. J., et al. (2014). Performance of an empirical bias‐correction of
Portals (GO‐ESSP). The CORDEX‐ a high‐resolution climate dataset. International Journal of Climatology, 34, 2189–2204.
AustralAsia data sets are downloaded Boé, J., Terray, L., Habets, F., & Martin, E. (2007). Statistical and dynamical downscaling of the seine basin climate for hydro‐
from https://esg‐dn1.nsc.liu.se/search/ meteorological studies. International journal of climatology, 27, 1643–1655.
esgf‐liu/. The observed daily gridded Burt, T. M. (2013). The silent river: The hydrological basis for river conservation. In S. Sabater, & A. Elosegi (Eds.), River conservation,
precipitation data were obtained from (pp. 39–60). Bilbao: Challenges and opportunities. Fundación BBVA.
the Bureau of Meteorology, Australia. Cannon, A. J. (2017). Multivariate quantile mapping bias correction: An n‐dimensional probability density function transform for climate
The data are available upon request via model simulations of multiple variables. Climate Dynamics, 1‐19.
http://www.bom.gov.au/climate/aust- Cannon, A. J., Sobie, S. R., & Murdock, T. Q. (2015). Bias correction of GCM precipitation by quantile mapping: How well do methods
maps/metadata‐daily‐rainfall.shtml. preserve changes in quantiles and extremes? Journal of Climate, 28, 6938–6959.
The streamflow station data are derived Charles, S. P., Bates, B. C., Smith, I. N., & Hughes, J. P. (2004). Statistical downscaling of daily precipitation from observed and modelled
from http://www.bom.gov.au/water/ atmospheric fields. Hydrological Processes,18(8), 1373–1394.
hrs/. The first author would like to Chen, J., Brissette, F. P., Chaumont, D., & Braun, M. (2013). Finding appropriate bias correction methods in downscaling precipitation for
acknowledge the financial support hydrologic impact studies over North America. Water Resources Research, 49, 4187–4205.
obtained from the Australian Awards Chen, J., Brissette, F. P., & Lucas‐picher, P. (2015). Assessing the limits of bias‐correcting climate model outputs for climate change impact
Scholarships (AAS). Authors also studies. Journal of Geophysical Research: Atmospheres, 120, 1123–1136.
gratefully acknowledge funding sup- Christensen, J. H., Boberg, F., Christensen, O. B., & Lucas‐Picher, P. (2008). On the need for bias correction of regional climate change
port from the Australian Research projections of temperature and precipitation. Geophysical Research Letters, 35, l20709.
Council (DP 180102737) for making Coron, L., Thirel, G., Delaigue, O., Perrin, C., & Aandréassian, V. (2017). The suite of lumped GR hydrological models in an r package.
this work possible. Environmental Modelling and Software, 94, 166–171.

NGUYEN ET AL. 19 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

Demaria, E. M. C., Palmer, R. N., & Roundy, J. K. (2016). Regional climate change projections of streamflow characteristics in the Northeast
and Midwest U.S. Journal of Hydrology: Regional Studies, 5, 309–323.
Dosio, A., & Paruolo, P. (2011). Bias correction of the ensembles high‐resolution climate change projections for use by impact models:
Evaluation on the present climate. Journal of geophysical research: atmospheres, 116, n/a–n/a.
Duan, Q. Y., Gupta, V. K., & Sorooshian, S. (1993). Shuffled complex evolution approach for effective and efficient global minimization.
Journal of Optimization Theory and Applications, 76, 501–521.
Eden, J. M., Widmann, M., Grawe, D., & Rast, S. (2012). Skill, correction, and downscaling of gcm‐simulated precipitation. Journal of
Climate, 25, 3970–3984.
Ehret, U., Zehe, E., Wulfmeyer, V., Warrach‐sagi, K., & Liebert, J. (2012). Hess opinions “should we apply bias correction to global and
regional climate model data?”. Hydrol. Earth Syst. Sci., 16, 3391–3404.
Ehsani, N., Vörösmarty, C. J., Fekete, B. M., & Stakhiv, E. Z. (2017). Reservoirs operations under climate change: Storage capacity options
to mitigate risk. J. Hydrol., 555, 435–446.
Evans, J. P. (2011). CORDEX—An international climate downscaling initiative. In MODSIM 2011 – 19th International Congress on
Modelling and Simulation – Sustaining Our Future: Understanding and Living withUncertainty, Modelling and Simulation Society,
Australian National University, Canberra, ACT, Australia (pp. 2705–2711). presented at MODSIM 2011–19th International Congress on
Modelling and Simulation‐Sustaining Our Future: Understanding and Living with Uncertainty, Perth, WA, 12 December 2011–16
December 2011.
Forzieri, G., Feyen, L., Rojas, R., Flörke, M., Wimmer, F., & Bianchi, A. (2014). Ensemble projections of future streamflow droughts in
Europe. Hydrol. Earth Syst. Sci., 18, 85–108.
Fowler, H. J., Blenkinsop, S., & Tebaldi, C. (2007). Linking climate change modelling to impacts studies: recent advances in downscaling
techniques for hydrological modelling. International Journal of Climatology, 27(12), 1547–1578.
Ghosh, S., & Mujumdar, P. P. (2008). Correction for bias in downscaling GCM simulations for hydrologic impact assessment. Proceedings of
Water Down Under, 2008, 440.
Giorgi, F., Jones, C., & Asrar, G. R. (2009). Addressing climate information needs at the regional level: the CORDEX framework. World
Meteorological Organization (WMO) Bulletin, 58(3), 175.
Graham, L. P., Andréasson, J., & Carlsson, B. (2007). Assessing climate change impacts on hydrology from an ensemble of regional climate
models, model scales and linking methods—A case study on the Lule river basin. Climatic Change, 81, 293–307.
Haerter, J. O., Hagemann, S., Moseley, C., & Piani, C. (2011). Climate model bias correction and the role of timescales. Hydrol. Earth Syst.
Sci., 15, 1065–1079.
Hagemann, S., Chen, C., Clark, D. B., Folwell, S., Gosling, S. N., Haddeland, I., et al. (2013). Climate change impact on available water
resources obtained using multiple global climate and hydrology models. Earth Syst. Dynam., 4, 129–144.
Hagemann, S., Chen, C., Haerter, J. O., Heinke, J., Gerten, D., & Piani, C. (2011). Impact of a statistical bias correction on the projected
hydrological changes obtained from three GCMs and two hydrology models. Journal of Hydrometeorology, 12, 556–578.
Hamon, W. R. (1961). Estimating potential evapotranspiration. Journal of the Hydraulics Division, 87, 107–120.
Hashimoto, T., Stedinger, J. R., & Loucks, D. P. (1982). Reliability, resiliency, and vulnerability criteria for water resource system perfor-
mance evaluation. Water Resour. Res., 18, 14–20. https://doi.org/10.1029/WR018i001p00014
Hay, L. E., Wilby, R. L., & Leavesley, G. H. (2000). A comparison of delta change and downscaled GCM scenarios for three mountainous
basins in the United State. Journal of the American Water Resources Association, 36, 387–397.
Johnson, F., & Sharma, A. (2011). Accounting for interannual variability: A comparison of options for water resources climate change
impact assessments. Water Resources Research, 47, w04508.
Johnson, F., & Sharma, A. (2012). A nesting model for bias correction of variability at multiple time scales in general circulation model
precipitation simulations. Water Resources Research, 48, w01504.
Johnson, F., & Sharma, A. (2015). What are the impacts of bias correction on future drought projections? Journal of Hydrology, 525,
472–485.
Lenderink, G., Buishand, A., & Van Deursen, W. (2007). Estimates of future discharges of the river rhine using two scenario methodologies:
Direct versus delta approach. Hydrol. Earth Syst. Sci., 11, 1145–1159.
Lu, J., Sun, G., McNulty, S., & Amatya, D. M. (2005). A comparison of six potential evapotranspiration methods for regional use in the
Southeastern United States. Journal of the American Water Resources Association (JAWRA), 41(3), 621–633.
Lv, X., Zuo, Z., Ni, Y., Sun, J., & Wang, H. (2019). The effects of climate and catchment characteristic change on streamflow in a typical
tributary of the Yellow River. Sci Rep, 9, 14535 (2019). https://doi.org/10.1038/s41598‐019‐51115‐x
Macadam, I., Argüeso, D., Evans, J. P., Liu, D. L., & Pitman, A. J. (2016). The effect of bias correction and climate model resolution on wheat
simulations forced with a regional climate model ensemble. International journal of climatology, 36, 4577–4591. https://doi.org/10.1002/
joc.4653
Maraun, D. (2016). Bias correcting climate change simulations—A critical review. Current Climate Change Reports, 2, 211–220.
Maraun, D., Shepherd, T. G., Widmann, M., Zappa, G., Walton, D., Gutierrez, J. M., et al. (2017). Towards process‐informed bias correction
of climate change simulations. Nature Clim. Change, 7, 664–773.
Mcmahon, T. A., Vogel, R. M., Pegram, G. G. S., Peel, M. C., & Etkin, D. (2007). Global streamflows—Part 2: Reservoir storage–yield
performance. Journal of Hydrology, 347, 260–271.
Mehrotra, R., & Sharma, A. (2006). A nonparametric stochastic downscaling framework for daily rainfall at multiple locations. Journal of
Geophysical Research: Atmospheres, 111, D15101. https://doi.org/10.1029/2005JD006637
Mehrotra, R., & Sharma, A. (2016). A multivariate quantile‐matching bias correction approach with auto‐ and cross‐dependence across
multiple time scales: Implications for downscaling. Journal of Climate, 29, 3519–3539.
Mehrotra, R., Sharma, A., & Cordery, I. (2004). Comparison of two approaches for downscaling synoptic atmospheric patterns to multisite
precipitation occurrence—art. no. D14107. Journal of Geophysical Research‐Atmospheres, 109, 14107–14107.
Mohammed, I. N., Bomblies, A., & Wemple, B. C. (2015). The use of CMIP5 data to simulate climate change impacts on flow regime within
the Lake Champlain basin. Journal of Hydrology: Regional Studies, 3, 160–186.
Nahar, J., Johnson, F., & Sharma, A. (2017). Assessing the extent of non‐stationary biases in GCMs. Journal of Hydrology, 549, 148–162.
Nguyen, H. (2018). A framework for correcting biases in climate model simulations in the frequency domain: Implications for hydrology.
PhD thesis, Civil and Environmental Engineering, Faculty of Engineering, UNSW, Australia. https://www.unsworks.unsw.edu.au/
permalink/f/5gm2j3/unsworks_52030
Nguyen, H., Mehrotra, R., & Sharma, A. (2019). Correcting systematic biases across multiple atmospheric variables in the frequency
domain. Climate Dynamics, 52, 1283–1298. https://doi.org/10.1007/s00382‐018‐4191‐6

NGUYEN ET AL. 20 of 21
19447973, 2020, 2, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019WR026022 by Cochrane Peru, Wiley Online Library on [08/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Water Resources Research 10.1029/2019WR026022

Nguyen, H., Mehrotra, R., & Sharma, A. (2016). Correcting for systematic biases in GCM simulations in the frequency domain. Journal of
Hydrology, 538, 117–126.
Nguyen, H., Mehrotra, R., & Sharma, A. (2018). Correcting systematic biases across multiple atmospheric variables in the frequency
domain. Climate Dynamics, 52(1‐2), 1283–1298.
Oudin, L., Hervieu, F., Michel, C., Perrin, C., Andréassian, V., Anctil, F., & Loumagne, C. (2005). Which potential evapotranspiration input
for a lumped rainfall–runoff model?: Part 2—Towards a simple and efficient potential evapotranspiration model for rainfall–runoff
modelling. Journal of Hydrology, 303(1‐4), 290–306.
Pagano, T., Hapuarachchi, P., & Wang, Q. (2010). Continuous rainfall‐runoff model comparison and short‐term daily streamflow forecast
skill evaluation. CSIRO Australia: Csiro Tech. Rep. EP103545.
Perrin, C., Michel, C., & Andréassian, V. (2003). Improvement of a parsimonious model for streamflow simulation. Journal of hydrology,
279, 275–289.
Piani, C., Haerter, J. O., & Coppola, E. (2010). Statistical bias correction for daily precipitation in regional climate models over Europe.
Theoretical and Applied Climatology, 99, 187–192.
Pierce, D. W., Cayan, D. R., Maurer, E. P., Abatzoglou, J. T., & Hegewisch, K. C. (2015). Improved bias correction techniques for
hydrological simulations of climate change. Journal of Hydrometeorology, 16, 2421–2442.
Rajczak, J., Kotlarski, S., & Schär, C. (2016). Does quantile mapping of simulated precipitation correct for biases in transition probabilities
and spell lengths? Journal of Climate, 29, 1605–1615.
Raupach MR, PR Briggs, V Haverd, EA King, M Paget, CM Trudinger (2009), Australian Water Availability Project (AWAP): CSIRO
Marine and Atmospheric Research Component: Final Report for Phase 3. CAWCR Technical Report No. 013. 67 pp.
Raupach, M. R., Briggs, P. R., Haverd, V., King, E. A., Paget, M., & Trudinger, C. M. (2012). Australian Water Availability Project. CSIRO
Marine and Atmospheric Research. Australia: Canberra.
Reshmidevi, T. V., Nagesh Kumar, D., Mehrotra, R., & Sharma, A. (2017). Estimation of the climate change impact on a catchment water
balance using an ensemble of GCMs. Journal of Hydrology, 556, 1192–1204.
Rocheta, E., Eevans, J. P., & Sharma, A. (2017). Can bias correction of regional climate model lateral boundary conditions improve low‐
frequency rainfall variability? Journal of Climate, 30, 9785–9806.
Sippel, S., Otto, F. E. L., Forkel, M., Allen, M. R., Guillod, B. P., Heimann, M., et al. (2016). A novel bias correction methodology for climate
impact simulations. Earth Syst. Dynam., 7, 71–88.
Sperna Weiland, F. C., Van Beek, L. P. H., Kwadijk, J. C. J., & Bierkens, M. F. P. (2010). The ability of a GCM‐forced hydrological model to
reproduce global discharge variability. Hydrol. Earth Syst. Sci., 14, 1595–1621.
Sperna Weiland, F. C., Van Beek, L. P. H., Kwadijk, J. C. J., & Bierkens, M. F. P. (2012). Global patterns of change in discharge regimes for
2100. Hydrol. Earth Syst. Sci., 16, 1047–1062.
Stehlı́k, J., & Bárdossy, A. (2002). Multivariate stochastic downscaling model for generating daily precipitation series based on atmospheric
circulation. Journal of Hydrology, 256(1), 120–141.
Teng, J., Potter, N. J., Chiew, F. H. S., Zhang, L., Wang, B., Vaze, J., & Evans, J. P. (2015). How does bias correction of regional climate
model precipitation affect modelled runoff? Hydrol. Earth Syst. Sci., 19, 711–728.
Teutschbein, C., & Seibert, J. (2012). Bias correction of regional climate model simulations for hydrological climate‐change impact studies:
Review and evaluation of different methods. Journal of Hydrology, 456–457, 12–29.
Themeßl, M. J., Gobiet, A., & Heinrich, G. (2012). Empirical‐statistical downscaling and error correction of regional climate models and its
impact on the climate change signal. Climatic Change, 112, 449–468.
Troin, M., Velázquez, J. A., Caya, D., & Brissette, F. (2015). Comparing statistical post‐processing of regional and global climate scenarios
for hydrological impacts assessment: A case study of two Canadian catchments. Journal of Hydrology, 520, 268–288.
Turner, S. W. D., & Galelli, S. (2016). Water supply sensitivity to climate change: An R package for implementing reservoir storage analysis
in global and regional impact studies. Environmental Modelling and Software, 76, 13–19.
Vrac, M., & Naveau, P. (2007). Stochastic downscaling of precipitation: From dry events to heavy rainfalls. Water Resources Research, 43(7).
Vrac, M., Noël, T., & Vautard, R. (2016). Bias correction of precipitation through singularity stochastic removal: Because occurrences
matter. Journal of Geophysical Research: Atmospheres, 121, 5237–5258.
Watanabe, S., Kanae, S., Seto, S., Yeh, P. J. F., Hirabayashi, Y., & Oki, T. (2012). Intercomparison of bias‐correction methods for monthly
temperature and precipitation simulated by multiple climate models. Journal of Geophysical Research: Atmospheres, 117, d23114.
Wilby, R. L., Charles, S., Zorita, E., Timbal, B., Whetton, P. and Mearns, L. (2004). Guidelines for use of climate scenarios developed from
statistical downscaling methods. Supporting material of the intergovernmental panel on climate change, available from the DDC of
IPCC TGCIA, 27.
Wilby, R. L., Hassan, H., & Hanaki, K. (1998). Statistical downscaling of hydrometeorological variables using general circulation model
output. Journal of Hydrology, 205(1), 1–19.
Wilby, R. L., Hay, L. E., Gutowski, W. J. Jr., Arritt, R. W., Takle, E. S., Pan, Z., et al. (2000). Hydrological responses to dynamically and
statistically downscaled climate model output. Geophysical Research Letters, 27(8), 1199–1202.
White, R. H., & Toumi, R. (2013). The limitations of bias correcting regional climate model inputs. Geophysical Research Letters, 40,
2907–2912.
Woldemeskel, F. M., Sharma, A., Sivakumar, B., & Mehrotra, R. (2014). A framework to quantify GCM uncertainties for use in impact
assessment studies. Journal of Hydrology, 519(part b), 1453–1465.
Woldemeskel, F. M., Sharma, A., Sivakumar, B., & Mehrotra, R. (2016). Quantification of precipitation and temperature uncertainties
simulated by CMIP3 and CMIP5 models. Journal of Geophysical Research: Atmospheres, 121, 3–17.
Wood, A. W., Leung, L. R., Sridhar, V., & Lettenmaier, D. P. (2004). Hydrologic implications of dynamical and statistical approaches to
downscaling climate model outputs. Climatic Change, 62, 189–216.
Zarch, M. A., Sivakumar, B., & Sharma, A. (2015). Assessment of global aridity change. Journal of Hydrology, 520, 300–313. https://doi.org/
10.1016/j.jhydrol.2014.11.033
Zhang, X. S., Amirthanathan, G. E., Bari, M. A., Laugesen, R. M., Shin, D., Kent, D. M., et al. (2016). How streamflow has changed across
Australia since the 1950s: Evidence from the network of hydrologic reference stations. Hydrol. Earth Syst. Sci., 20, 3947–3965.

NGUYEN ET AL. 21 of 21

You might also like