Amplification of risks to water supply at 1.5 °C and 2 °C in drying climates: a case study for Melbourne, Australia

Rainfall and temperature observations are from the Australian Bureau of Meteorology's Australian Water Availability Project (AWAP) (Jones et al 2009). Inflow observations for Melbourne's four major water harvesting reservoirs (Thomson, Upper Yarra, O'Shannassy and Maroondah) are supplied by Melbourne Water. Climate model data are from the CMIP5 (Taylor et al 2012a) archive.

To assess the risk of changes in precipitation (P) and temperature (T) to a water supply system we combine a stochastic hydrological risk assessment with climate model projections of future rainfall and temperature change (figure S1 shows the overall methodology and is available online at stacks.iop.org/ERL/14/084028/mmedia). We start by synthetically generating 1000 replicates of monthly P and T of the same length as the observed series, 100 years. This large number of replicates from the stochastic approach allows for more precise risk estimates than are possible if only one replicate (e.g. the observed series) is used. This is because risks computed from a stochastic ensemble are less sensitive to sampling variability. With regards to model accuracy, the stochastic model accounts for both the correlation between P and T and the temporal autocorrelation of both series, using a multi-site lag-one autoregressive AR(1) model with Box-Cox transformation (Matalas (1967) and section 2.3 in the study by Peel et al (2015) contain detailed explanations). The model preserves the annual probability distributions (figures S3–4), annual serial autocorrelation functions for lags of 1–20 years (figures S5–6) and the covariance structure (figure S7) of water year (March–February) P and T. We determine therefore that the model is a suitably accurate representation of rainfall and temperature at this spatio-temporal scale. We then simulate mean-state hydroclimatic change using annual shifts, ΔP (%) and ΔT (°C), from a 1961–1990 baseline. We apply these shifts on a regular grid (ΔP: −25 to +25%, ΔT: −0.5 to +3.0°C) to our synthetic water-year P and T time series. We then disaggregate our synthetic time series from annual to monthly time step using a non-parametric k-nearest neighbour (kNN) bootstrap sampling approach (Lall and Sharma 1996). In this method, monthly fragments of the stochastically generated annual total are sampled probabilistically from observed annual data (Henley et al 2013) to obtain realistic sequences of monthly P and T that resemble the annual and interannual characteristics of the observed data.

We then generate monthly reservoir inflow sequences, Q, using the lumped conceptual Precipitation-Evapotranspiration-Runoff Model (PERM) of Peel et al 2015, (see section 2.5 of that study). The inputs to PERM are monthly P and T. The output is monthly catchment runoff, Q, which we take as total inflow into the system's four major water storages. A time series of T is used as an input to PERM rather than potential evapotranspiration (PET), as climate model projections of T are considered more reliable than projections of variables required to estimate PET. In the case study region, this decision is likely to be acceptable as the catchments are largely water-limited (McMahon et al 2015), however this may not be true for other types of catchments. The model is calibrated using P, T and Q data from 1913–2012. The Nash–Sutcliffe Efficiency (NSE) for monthly runoff is 0.82, which indicates a good calibration.

Reservoir behaviour is simulated with a mass balance model similar to that used by Henley et al (2013). The model includes a seasonally-varying monthly demand, increasing demand due to population growth, savings from water restrictions, system losses and augmented supply from desalination. The reservoir's initial storage level is simulated probabilistically with a gamma distribution (figure S8). Water demand and losses are modelled with a seasonally varying pattern linked to monthly mean temperature (figure S9). The demand pattern is assumed to remain constant in a warmer climate. To test the sensitivity of increasing demand, we represent higher population demand scenarios by increasing the total demand (D) as a fraction of the historical annual average inflow (Q) into the four major water storages (594 GL) from a 'low' demand level of D/Q = 0.7 to a 'very high' demand level of D/Q = 0.9. The highest level simulated here is an approximation of the median demand projection at around 2045–2050 (City West Water et al 2017). Water restrictions and other demand management measures, which may include water pricing, are simulated in four stages such that the total demand reduces by 5%, 10%, 20% and 25% when the total reservoir storage level falls below 60%, 45%, 20% and 10%, respectively. Note that managed aquifer recharge is not possible as a major water supply measure in the study region. The water supply shortage risk is calculated by dividing the number of months spent under the severe shortage limit (10% storage) by the total number of months in each replicate, and averaging this probability across all 1000 stochastic replicates.

Climate model simulations were obtained from the fifth phase of the Climate Model Intercomparison Project (CMIP5). Simulations from 15 models in the CMIP5 archive (listed in table S1) over the period 1861–2005 under both natural and anthropogenic forcings (historical), 1861–2005 under natural forcings only (historicalNat), and 2006–2100 under four representative concentration pathway (RCP) scenarios (RCP2.6, RCP4.5, RCP6.0 and RCP8.5) were utilised. Modelled fields are re-gridded to a common 2° x 2° grid. We use a time-slice method (James et al 2017, King et al 2017) to select model-simulated years from the four RCP scenarios to represent the 'Natural', 'Current', '1.5 °C' and '2.0 °C' warmer worlds. The 'Natural' scenario is estimated using the 1901–2005 period from the ensemble of the historicalNat simulations. These are simulations in CMIP5 that include natural external climate forcings such as solar variability and volcanic emissions but exclude anthropogenic climate forcings. The 'Current' scenario is defined as the 2006–2026 period (that is, a 21 year window centred on 2016) in the high greenhouse gas emissions RCP8.5 scenario. We define our 1.5 °C world as all years within decades in the 2006–2100 experiments in which decadal-average global temperatures are between 1.3 °C–1.7 °C warmer than the mean global temperature in the equivalent model's historicalNat simulation for 1901–2005. The 2 °C world is defined similarly but using a decadal-averaged temperature of 1.8 °C–2.2 °C above the corresponding natural baseline. We use each CMIP5 model's mean change in precipitation (in %) and temperature (in °C) relative to its own baseline, implicitly bias-correcting the mean. For our case study we use two 2° × 2° grid cells over the study region.

Several highly populated regions of the world face the prospect of concurrent precipitation decrease and rising temperatures under global warming (figure 1(a)). Observations and model projections indicate the American Southwest is shifting towards increasing aridity (Seager et al 2007, Cook et al 2015) (figure 1(b)). Although some of the temperature increase during meteorological drought is the result of land-surface feedbacks from reduced evaporative cooling and increased incoming solar radiation from reduced cloud cover (Wolf et al 2017), anthropogenic warming is increasing the probability of concurrently warm and dry conditions resembling those during the severe 2012–2014 drought in California (Diffenbaugh et al 2015). A broad and consistent drying signal is projected in the Mediterranean (figure 1(c)), with drought amplification projected under increasing levels of global warming (Marx et al 2017, Samaniego et al 2018). Southern Australia (figure 1(d)) is a region experiencing strong influences from both climate change and interannual to decadal variability (Cai et al 2014). South-western Australia has experienced severe declines in rainfall and runoff since the 1970s, driven by anthropogenic emissions of greenhouse gases (Delworth and Zeng 2014), stratospheric ozone depletion (Delworth and Zeng 2014) and regional land-use change (Pitman et al 2004).

Zoom In Zoom Out Reset image size

Here we compare mean precipitation and temperature shifts in the American Southwest, the Mediterranean and Southern Australia for four global warming scenarios: a pre-industrial 'natural' world, the 'current' world, and two future worlds where global mean surface temperature is 1.5 and 2.0 °C above natural baseline temperatures (King et al 2017). We utilise a suite of model experiments from the fifth phase of the Coupled Model Intercomparison project (CMIP5) (Taylor et al 2012b) to obtain large ensembles of simulated climate at these levels of global warming (King et al 2017, Methods). In the three regions, models generally project concurrent warming and drying. Mean temperature change (ΔT) follows the general trajectory of global temperature, with higher rates of projected regional warming in the Mediterranean and in the American Southwest than in southern Australia (King et al 2017) (figures 1(b)–(d)). At 1.5 °C of warming all three regions are projected to experience precipitation reductions relative to 1961–1990 (figures 1(b)–(d)). In the Mediterranean the mean precipitation change at 1.5 °C is –5.3%, with 90% confidence intervals of [–12.4%, +0.8%]. For Southern Australia the projection is for a change of –2.6% [–9.6%, +2.4%], and for the American Southwest it is –2.2% [–6.4%, +1.5%].

Australia felt the consequences of extremely dry and warm conditions during the Millennium Drought from 1997–2009 (figures 2(d)–(g)). The drought led to the commissioning of several large water supply augmentation projects, including for Melbourne. The drought was associated with substantial agricultural losses and environmental degradation (van Dijk et al 2013). The drought was unprecedented in instrumental records in southeastern Australia, and it is plausible that climate change had an influence (van Dijk et al 2013, Cai et al 2014, Hope et al 2017). For the city of Melbourne's major surface water supply catchments (Rhodes et al 2012), the Millennium Drought brought declines in reservoir inflow in all months (figures 2(a) and (b)), with the highest volume reductions in winter and spring. Monthly inflow was between 22%–55% below average (figure 2(c)). The highest proportional runoff reductions occurred during autumn. For temperature, the positive anomalies were strongest in summer. At a given monthly rainfall total, higher mean monthly temperatures are coincident with reductions in catchment runoff. This relationship is strongest from Sep–May.

Zoom In Zoom Out Reset image size

The city of Melbourne demonstrated a capability to adapt to water shortages during the Millennium Drought (Rhodes et al 2012). However, the possibility of droughts more intense than previously experienced (Tan and Rhodes 2013), higher temperatures (DELWP 2016a), uncertainty in catchment responses (Saft et al 2016, Tan and Neal 2018) and demand growth (Melbourne Water 2017) are all major reasons for concern. Future augmentation of the desalination plant may mitigate these concerns but may come at a substantial cost to the community and the environment. Note that the State Government has offset the current operational power of the desalination plant with renewable energy certificates (DELWP 2019). The Millennium Drought left water managers in southern Australia confronted with a pressing need to better understand the future likelihood of severe and protracted droughts.

In this study we evaluate future risks to water supply infrastructure at particular levels of global warming both with, and without, the addition of water supply from a desalination plant. We use climate models and hydrological modelling to infer future drought risks at key levels of global warming. To demonstrate the method, we focus on the broad characteristics of Melbourne's major surface water supply. The population of greater Melbourne is projected to grow from 4.5 million in 2015 to around 8.0 million in 2051 (DELWP 2016b). With rapid population growth, projected warming and drying, and long instrumental observations (>100 years), the city's water supply system is an ideal case study. The recently observed severe drought and the possible influence of climate change provide additional impetus to understand the combined impacts of future changes on risks to water supply at the catchment scale at natural, current and policy-relevant future levels of global warming.

Here we assess the probability of severe water supply shortage for a range of shifts in mean precipitation and temperature. Using hydrological modelling we simulate total inflows into a water supply reservoir system and subsequent storage levels using stochastic precipitation and temperature at a monthly time step and conceptual rainfall-runoff modelling (Methods). The PERM rainfall-runoff model (Peel et al 2015) represents interception by vegetation, soil moisture storage and evapotranspiration that is limited by either soil moisture or energy. Our simulations use a simplified representation of the major water supply catchments and reservoir system for Melbourne, both with and without a climate independent water source—a desalination plant capable of supplying up to 150 GL of water per annum. The models capture the key statistical properties of the climate variables. In capturing the interannual persistence, annual probability distributions and the covariance between T and P (see figures S3–7) the models simulate realistic hot-dry sequences. We incorporate plausible water restrictions policies and savings that are consistent with projected drought response actions (Methods).

Mean precipitation reduction is the primary driver of increasing water shortage probability (contoured surfaces in figure 3), producing a rapid escalation of the probability that storage levels are below 10%, a level considered here to indicate severe supply shortages. Water supply shortage risk increases as a function of both precipitation decline and temperature increase. Relatively small changes in both of these variables are capable of substantially increasing the risk. Accounting for the negative correlation between catchment temperature and precipitation, higher mean temperatures can exacerbate risk by increasing the moisture demand from the atmosphere, increasing evapotranspiration from the catchment and consequently reducing the inflow into the reservoir system. For example, with a mean catchment warming of +2 °C relative to 1961–90 and a precipitation change of –5%, the water supply shortage risk intensifies to over 30% in the absence of the desalination source (figure 3(a)). The addition of a major desalination source (figure 3(b)) substantially reduces the risk.

Zoom In Zoom Out Reset image size

Climate model projections of mean precipitation and temperature change (ΔP and ΔT) over the Melbourne water supply catchment region are now superimposed onto the risk surface (markers on figure 3) for each of the four warming levels: Natural, Current, 1.5 °C and 2.0 °C of global warming. The CMIP5 model projections indicate that regional warming will occur alongside global warming. There is high variability among the climate models. However, the models on average project a reduction in precipitation at both the 1.5 °C and 2 °C levels of global warming for the case study region. The mean precipitation shift over the catchments is –2.0%, with 90th percentile limits of [–11.0%, +5.1%] relative to 1961–90 at the 1.5 °C warming level. At 2 °C global warming the shifts are –3.3% [−14.7%, +3.3%]. The point on these risk surfaces directly below each marker can then be used to infer the corresponding risk. As the world warms, many regions will experience increasing water demand due to population growth. Melbourne's population is expected to nearly double over the coming three decades. We therefore compare the water supply shortage risk for five possible water demand scenarios under the four global warming levels, both without (figure 4(a)) and with desalination (figure 4(b)). As future demand on the reservoir system increases from 0.7 to 0.9 of the observed annual average inflow, the severe water supply shortage risk at 2 °C of global warming increases rapidly in the absence of desalination. At an annual demand of 0.75 of the mean inflow, the ΔP and ΔT climate model projections for the Melbourne region infer a median water supply shortage risk of only 0.6%, with 90th percentile limits of [0.1%, 4.3%] at 1.5 °C of warming. This increases to 2.9% [0.1%, 12.1%] at 2.0 °C of warming in the absence of system augmentation. At the higher demand level of 0.85 of the mean annual inflow, the median water supply shortage risk is 9.6% [3.8%, 24.4%] at 1.5 °C, increasing substantially to 20.4% [3.9%, 39.5%] at 2 °C of global warming. With desalination, the risks are substantially ameliorated (figure 4(b), note y-axis limit change). At the highest demand level however, the median risk increases rapidly with the degree of warming.

Zoom In Zoom Out Reset image size