Sensitivity of Future U.S. Water Shortages to Socioeconomic and Climate Drivers: A Case Study in Georgia Using an Integrated Human-Earth System Modeling Framework
Climatic Change
Michael J. Scotta,*, Don S. Dalya, Mohamad I. Hejazib, G. Page Kyleb, Lu Liub, Haewon C. McJeonb, Anupriya Mundrab, Pralit L. Patelb, Jennie S. Ricea, Nathalie Voisinc
aPacific Northwest National Laboratory, Richland, Washington, USA
b Joint Global Change Research Institute, Pacific Northwest National Laboratory/University of Maryland, College Park, Maryland, USA
cPacific Northwest National Laboratory, Battelle Seattle Research Center, Seattle, Washington, USA
*Corresponding Author: Tel: +1 2065283312; fax +1 509 372 6710; E-mail address:(N. Voisin);Mailing address: 1100 Dexter Avenue N, Suite 400, Seattle, WA, USA 98109
Supplemental Material
S.1 Selected Additional References on Water Demand and Supply Modeling Not Included in Main Text (Citations in Reference List for Supplemental Material)
Reference and Year / Geography:Single Basin (SB), Multiple Basins (MB), Multi-Region, Including World (MR), / Water Source (s):
Surface (S), Ground (G) / Modeled Components:
Demand (D), Supply Only (S), Both (DS), Water Management (WM) / Type of Uncertainty/Sensitivity Assessment:
Multiple Socioeconomic Demand Scenarios (MD), Multiple Climate Scenarios (MC), Multiple Climate Models (MM), Multiple Downscaling Methods (MDown), Multiple Hydrologic and SectorModels/Methods (MS), Monte Carlo Uncertainty Propagation (MU)
Alcamo et al. 2007 / MR,MB / S / DS / MD, MC, MM
Arnell et al. 2003 / MR, MB / S / S / MC, MM
Averyt et al. 2011 / MB / S / DS
Blanc et al. 2014 / MB / S,G / DS,WM / MD,MC, MM
Brown et al. 2013 / SB / S / S, WM / MC, MM
Byers et al. 2014 / MB / S / DS, WM / MD
Chen et al. 2014 / SB / S / S / MC, MD
Crosbie 2013a / SB / G / S / MC, MM
Crosbie 2013b / MB / G / S / MC, MM
de Graaf et al. 2015 / MR / G / S / MU (no climate forecast)
Döll 2009 / MR / G / S / MC,MM
Ekstrom et al. 2013 / SB / S, G / S, WM / MC,MM
Elgaali et al. 2007 / SB / S / D,WM / MD, MC,MM
Elliott et al. 2014 / MR / S, plus renewable G / DS, WM / MD,MC,MM,MS
Ghosh and Katar 2012 / SB / S / S / MC, MM,MDown
Gu et al. 2015 / SB / S / S
Hanasaki et al. 2013a, 2103b / MR / S / DS,WM / MD,MC,MM
Leng et al 2015 / MR / S
Matonse et al. 2013, / SB / S / DS
Ouyang et al. 2014 / SB / S / S / MDown
Parkinson and Djilali 2015 / MB / S / DS / MD, MC
Portmann et al. 2013 / MR / G / S / MC, MM
Rasmussen et al. 2014 / SB / S / S / MC
Roy et al. 2012 / MB / S / DS / MC,MM
Schewe et al. 2014 / MR / S / D, S / MC,MM, MS
Singh et al. 2015 / MB / S
Strzepek et al. 1995 / MB / S
Tang and Lettenmaier 2012 / MR / S / S / MC, MM
Vano et al. 2010 / SB / S / DS / MC, MM
van Vliet et al. 2013 / MB / S / S / MD, MC, MM
Vörösmarty et al. 2000 / MR / S
Wada et al. 2014 / MR / (unconstrained) / D / MC, MM, MS
Yoshikawa et al. 2014 / MR / S / DS, WM / MM, MS
S.2 Additional Information on the Flint River Drainage
The Flint River, with two run-of–the-river dams, is a largely unmanaged river (Saldirriaga 2010). It joins with the Chattahoochee River, a heavily controlled and managed river (16 total dams, 5 large dams, 3 large reservoirs), to become the Apalachicola River, which drains into the Gulf of Mexico (2015 Chattahoochee River Dams, Lownsbery 2014). The three rivers together comprise the Apalachicola-Chattahoochee-Flint (ACF) river basin, the management of which has been the subject of much contention among the states of Georgia, Alabama, and Florida Wright et al. 2012). As the result of historical episodes of drought and disputes over water priorities, the ACF basin has had a series of water studies and management plans designed to allocate water during drought (EPD 2006, Upper Flint Council 2011a,Lower Flint-Ochlockonee Council 2011a, Wright et al. 2012). Although some of the analyses include wet and dry climates, none of these plans as yet addressess climate change. About 10 percent of the water withdrawn in the basin is used in municipalities and industry, and 90 percent is used in agriculture, predominantly in irrigation. Sixty-six percent of irrigated acreage is irrigated with groundwater, but the supplying aquifers (Claiborne, Cretaceous, and Upper Floridian) are closely connected to the surface water, so that ground water withdrawal reduces surface flow, limiting the use of groundwater as a substitute to protect surface water minimum flows (Wright et al. 2012). Drought protection actions have included auctions for compensation to suspend irrigation. In addition, specific technical and behavioral conservation actions on irrigation systems have reduced irrigation water demand. Future considerations include further irrigation efficiency upgrades to further limit withdrawals per unit of consumptive demand as wellas additional storage (UpperFlint Council 2011b, Lower Flint-Ochlockonee Council 2011b). Independently, the ACF basin also has attracted attention of researchers concerned with the interacting effects of climate change and human water demand on water availability (Lettenmaier et al. 1999, Georgakakos et al. 2010, Lownsbery 2014). Though different in their methods and assessments of future water availability, all three basin-specific studies found that projected increasing demand would put additional pressure on water system performance in the context of climate change and both Georgakakos and Lownsbery found that lower water supplies under climate change also would do so.
Although those analyses were much more explicitly designed for this basin than the large scale modeling exercise in this paper, they have results that can be compared qualitatively with our results. In Lettenmaier et al. (1999), which only dealt with surface water, there was an overall increase in runoffinto the system in the 2050s due to increased precipitation, but water management results were as sensitive to demand growth and water system operations as to climate, similar to our findings. The hydrological model in Georgakakos et al. (2010) included groundwater recharge but not the strong surface water-groundwater interaction in the lower Flint region which generally increases surface flow. The future climate was wetter than historical in about 20% of the future scenarios and drier in the rest, with both greater risk of flooding and more extreme droughts. Future mean runoff was generally lower than historical values. ACF gross water withdrawals increased from 2007 to 2050, but net withdrawals decreased in the Flint part of the basin. Water management findings included: a) reduced reservoir levels, b) lower electricity production, and c) greater inflow water deficits with climate change in 5 out of 13 climate models in the A1B and in 10 out of 13 models in the A2 emissions scenarios. The average amount of shortage was doubled by projected growth in water demand in the A1b scenario and increased by 58 percent in the A2 scenario. Our results are also sensitive to future demand. In Lownsbery (2014), demand scenarios as well as supply scenarios were stochastic and were combined into 13,500 combinations of water demand and climate changes. Floridian formation groundwater was available to serve demand in the lower Flint River, and groundwater interactions were included, with 10 gallons of groundwater equivalent to 6 gallons of surface water. This would reduce the impacts of agricultural demand on water stress relative to the current paper, where all demand is assumed to be served by surface water. The climate projections were based on 95 ensemble members from Climate Model Intercomparison Project Phase 3 (CMIP3) and 150 scenarios from CMIP5, bias-corrected and statistically downscaled to the ACF basin. Water demand in Lownsbery’s study was not affected by climate. Results included calculations of “reliability” (probability of a water deficit at Atlanta, Columbus, and Blountstown) and “vulnerability” (average size of water deficits, measured as withdrawals not consumption at the same locations). Vulnerability is closer to the deficit metric used in the current study. Unfortunately for comparison purposes, none of the reported locations in Lownsbery isolates the Flint River, where much of the irrigation is located, vulnerability appears to be defined in terms of withdrawals rather than consumption, and groundwater serves agricultural demand in particular. Thus, while future growth municipal and industrial demand in Lownsbery (much of which is projected for the Atlanta area) increases mean vulnerability by about 5 to 20 percent and climate increases vulnerability by about 8 percent, the impact of agriculture demand is less—an increase of about 2 to 5 percent.
Figure S-1illustrates the modeled average monthly flows of the Flint River at Bainbridge and Chattahoochee River at Columbus from the CLM-MOSART-WM modeling, compared with observed 1985-2004 monthly averages. The modeled historical flows, though not from a watershed-specific operations model, closely match observations for the Flint River, which is largely unregulated. The Chattahoochee has five major dams, three large reservoirs, and is used to provide hydropower, flood control, navigation, recreation, water quality, and wildlife protection in addition to the various consumptive uses modeled for this paper (Lownsbery 2014). Even though these are largely non-consumptive demands, they do alter river flows. This analysis does not model these non-consumptive demands and does not match the Chattahoochee flows as well as those of the Flint. The figure also shows the reference case modeled increase in average monthly flows increase for the 2040-2059 period for the RCP8.5 climate.
(a)
(b)
Fig S-1. PRIMA modeled average monthly flows of (a) the Flint River at Bainbridge, GA and (b) the Chattahoochee River at Columbus, GA, compared with observed values for 1985-2004 (Historical) and for 2040-2059 with RCP 8.5 climate.
Table 1 in the main text demonstrates that a structured sensitivity analysis can be used to show when a given policy might be effective and when it might not. For example, if Georgia extended its existing program of technological assistance and financial incentives to irrigators, the state would be able to significantly reduce growth in irrigation water demand (Upper Flint Council 2011a, Lower Flint-Ochlockonee Council 2011a). In the table a target 25percent reduction in growth of irrigation consumption in the Flint River sub-basin would eliminate the surface water deficit in the low and reference cases and reduce but not eliminate the surface water deficit in the highest demand scenario. This could be a feasible target. Only about half of Flint River irrigation systems in the region in 2011 were believed to be efficient low-pressure or drip systems, only about half of growers performed the improved water scheduling that would reduce consumptive use, and both these percentages could be raised substantially ((Upper Flint Council 2011b, Lower Flint-Ochlockonee Council 2011b). Also, over the last few years, by using a number of technical “fixes” such as metering and better scheduling, irrigators in Georgia have been able to substantially reduce their water usage.
S.3 Additional Detail on Water Demand,Factorial Analysis, and Coefficients
Eight groups of variables (“factors”) affect water consumption in the PRIMA framework. Population and Gross Domestic Product (GDP) per worker are two factors that determine the scale of activity in all water demand sectors. State, national, and world population and GDP follow GCAM model scenarios consistent with RCP8.5. The six major water demand sectors are domestic, manufacturing, primary energy, livestock, electricity generation, and agriculture (crops). Each water demand sector, has a specific set of variables (demand coefficients) governing the rate of end use water consumption per capita or per unit of activity. The six sector-specific groups of demand coefficients can take on a range of values reflecting parameter uncertainty in modeled demand for water. Each individual sector’s set of demand coefficients is treated together in this paper as a single factor.
Water demand for both irrigation and non-irrigation use is affected by human population and economic activity (Hejazi 2014). In the PRIMA modeling system demand for agricultural products is felt through demand for food and non-food products and the land used to grow them. The major Flint River sub-basin crops like wheat, corn, soybeans, and cotton are traded globally. National and international population and economic production (gross domestic product or GDP) affect the scale of demand of each agricultural product in each area where it is grown and therefore the proportion of land and amount of water needed to grow it. Most other sources of demand for water are associated with state and local population and economic production. Scenarios of national and state population were used to span a significant range of population uncertainty (Scott et al. 2015).
S.3.1 Population and GDP
Georgia population grew from 6.5 million in 1990 to 9.9 million in 2012(Census Bureau 2012). By 2050, population was assumed in this study to grow to between 12.1 million and 14.9 million, based on Scott et al. (2015). Future population scenarios were constructed from long-range low, reference, and high U.S. Census Bureau and U.N. population forecasts for the United States, combined with trended state population shares constructed from U.S. Census Bureau state-level and national forecasts. Median and range of average annual rates of population increase from 2010 to 2095 were reference (median), 1.6 percent per year; range, 1.3 to 1.8 percent. Details are available in Scott et al. (2015).
Georgia GDP per worker (a measure of economic productivity and wealth) grew from $62,986 in 1997 to $70,738 in 2011 (2005 chained dollars), an annual rate of growth of 0.8 percent per year for a 14-year period that included both an economic boom and a serious recession. Future trends in GDP per worker were estimated by statistically sampling from a joint distribution of 28 correlated historical 15-year Georgia and national GDP trends between 1970 and 2004, and applying sampled values to the forecast periods. Output per worker is projected to grow at between 1.19 percent and 1.27 percent per year between 2010 and 2050. See Scott et al. (2015) for details of the projection methodology.
S.3.2 Factorial Analysis of Demand Coefficients
We characterized the uncertainty in the downscaled water demands from GCAM-USA using a one-third fractional factorial design, described below, that feature 8 “factors” or groups of variables that affect water demand. These werepopulation, gross domestic product(GDP) per worker (a proxy for the income level of the population based on productivity of the workforce), which affect all sectors, and six groups of sector-specific economic and technological input variables that control water consumption in GCAM’s six water-consuming sectors, shown in Table S.1. Each sector’s input variables affects water demands, each of which can take literature-based values. Each level was a fixed combination of those inputs corresponding to their low, reference,and high settings. Each level is a point in the joint distribution of the subset of inputs underpinning a factor, with the reference level combination sampling the reference level of the underpinning joint distribution and the low and high level combinations We characterized the variability in water demand with respect to variability in the GCAM inputs with a fractional factorial computer experiment (Montgomery 2012). A full factorial computer experiment provides data to infer the effect of each factor (inputs), say A, B, C or D, or statistical interaction between factors, say AB, ABC or ABCD, on the response (output) using Analysis of Variance (ANOVA). A fractional factorial experiment allows us to make inferences about some but not all factors and factor interactions with the fraction usually selected to estimate main effects (A, B, C, D) and lower-order interactions (AB, BC…) but not all higher-order interactions (ABCD). Unless the response surface of the model is highly irregular in shape, including lower-order interactions is usually sufficient to capture a high proportion of total variance in the output of the model. GCAM’s response surface fits this description (Scott et al. 2014).
Here, we used a fractional factorial design to systematically sample the distribution of GCAM inputs in order to make inferences about the water demand distribution from the resulting GCAM-calculated water demands. In particular, we included a subset of factor combinations representing the extremes of the GCAM inputs in order to estimate the low and high water demands.
In this experiment, a factor was one of 2 individual and 6 composite inputs that affected water demand (a composite input is a group of related GCAM inputs that were varied together as one input). We identified the literature-based low, mid and high values for each input and associated these with the L, M and H levels of the corresponding factor.
The combination of levels for each of the 8 factors, say HHLMLLHM, defined one parameterization of GCAM. We generated a 38-1 fractional factorial design featuring 2,187 of the 6,561 possible factor combinations. We identified the one-third fraction using the defining relation I = ABCDEFGH; i.e., theeight factors in various combination (Montgomery 2012). Our selected fraction was chosen to reveal extreme water demands with a known provenance.We observed that the mid and extreme water demands corresponded directly to the mid and extreme combinations of inputs – LLLLLLLL, MMMMMMMM, and HHHHHHHH.These runs representing the observed central tendencies and extremes of GCAM-predicted water demand were then downscaled for further analysis.
For each water demand-side parameter uncertainty in GCAM-USA, we developed low and high scenarios using one of three different methods: (a) based on a detailed literature assessment of range, (b) based on aggregate estimates from other studies, and (c) based on halving and doubling of the GCAM-USA reference case parameter values when (a) or (b) were unavailable. We then selected the reference demand and extreme low and high water demand results from the fractional factorial analysis for the South Atlantic-Gulf Basin in the period 2040-2059 to use in the downscaling for the demand-side inputs to the WM model runs. Figure S.2shows the distributions of annual water use demand by end use with their highest and lowest value traces (H, L) at their native geographical level in GCAM prior to temporal and geographical downscaling.
Figure S-2. Annual consumptive water demand by end use prior to geographic downscaling. Red and green traces are the highest and lowest demand traces for all factors set to maximize and minimize demand in the GCAM model socioeconomic scenario corresponding to RCP8.5 greenhouse gas concentrations.
Table S.1 shows the water consumption coefficients used in this study. The water use efficiency and income and price elasticity values for the domestic water sector and for all three scenarios are taken from Hejazi et al.(2013). The reference value for the technology improvement rate came from Hejazi et al. (2013) as well; however, the low and high values were created by halving and doubling the reference value respectively. The reference water use coefficients for the manufacturing, primary energy, and livestock sectors were adopted from Hejazi et al. (2014). The low and high values were generated by reconciling estimates from Hejazi et al. (2013) and Vassolo and Döll (2005) for the manufacturing sector, and Hejazi et al. (2014) and Steinfeld et al. (2006) for livestock. For primary energy, the low and high values were calculated by halving and doubling of the reference value, respectively. All the coefficient values for electricity generation for all three scenarios were taken from Davies et al. (2013). The irrigation water coefficients were taken from Chaturvedi et al. (2013) and the references therein.