Water Resources Adaptation to Climate and Demand Change in the Potomac River

8 The effects of climate change are increasingly considered in conjunction with changes in water 9 demand and reservoir sedimentation in forecasts of water supply vulnerability. Here, the relative 10 effects of these factors are evaluated for the Washington, DC metropolitan area water supply for the 11 near (2010 to 2039), intermediate (2040-2069), and distant future (2070 to 2099) by repeated water 12 resources model simulations. This system poses water management challenges due to long water 13 delivery travel times that increase uncertainty, multiple water jurisdictions that constrain potential 14 decisions, and future scenarios that simultaneously increase demand and decrease water supply 15 during the critical summer period. Adaptation strategies were developed for the system using a 16 multi-objective evolutionary algorithm. Optimized reservoir management policies were compared 17 using six distinct objectives, ranging from reservoir storage to environmental and recreational 18 benefits. Simulations of future conditions show water stress increasing with time. Reservoir 19 sedimentation is projected to more than double (114% increase) the severity of reservoir storage 20 failures by 2040. Increases in water demand and climate change are projected to further stress the 21 system, causing longer periods of low flow and a loss of recreational reservoir storage. The adoption 22 of optimized rules mitigates some of these effects, most notably returning simulations of 207023 2099 climate to near historical levels. Modifying the balance between upstream and downstream 24 1 Stagge, April 23, 2017 reservoirs improved storage penalties by 20.7% and flowby penalties by 50%. Changing triggers 25 for shifting load to off-line reservoirs improved flowby (8.3%) and environmental (4.1%) penalties 26 slightly, while changing demand restriction triggers provided only moderate improvements, but 27 with little adverse effects. 28


Introduction
Climate research indicates that the Earth's climate is changing in response to changes in the global atmospheric composition, brought about by human activities (IPCC 2014). With atmospheric research improving the reliability of climate projections, water resources planners and engineers must consider climatic changes as important factors for water supply planning, along with more traditional nonstationary factors such as demand change and reservoir sedimentation. Once future vulnerabilities to any of these factors are identified, adaptation strategies can be developed to mitigate their effects. Like many major cities, the Washington, DC metropolitan area (WMA) is interested in identifying changes in water supply vulnerability arising from (1) increased water demand, (2) losses of storage, and (3) changes in natural water availability because of the effects of climate change. This study explores these questions and demonstrates how water resources optimization can be combined with projections of future conditions to develop adaptation strategies using the WMA as a case study.
The WMA is the sixth largest metropolitan area in the United States (U.S. Census Bureau 2016), housing an estimated 6.1 million residents across 15 counties in Maryland (MD), Virginia (VA), and the District of Columbia (DC). Each of these three regions operate under separate water suppliers, creating an interesting jurisdictional challenge that was largely addressed by a unique shared decisionmaking scheme designed to ensure equitable water access during water shortages (U.S. Army Corps of Engineers 1982). Water for the region (Fig. 1) is primarily provided by withdrawals from the Potomac River, whose flow can be augmented by the Jennings Randolph Reservoir, located a 9-10 day travel time (300 km) upstream of the Washington, DC water supply intakes, and the smaller Little Seneca Reservoir, located only a 1-day travel time upstream, which can be used to fine-tune releases (Sheer and Flynn 1983). This design, completed in 1982, allows the 38,000 km 2 Potomac watershed to remain largely uncontrolled, but also increases the importance of effective water management policies. Maryland and Virginia maintain off-line water storage, the Patuxent and Occoquan Reservoirs, respectively, which can supplement water extracted from the Potomac River. In 2008, 31% of suburban Maryland's water production came from the Patuxent Reservoir and 42% of suburban Virginia's water production came from the Occoquan Reservoir, with the remainder and all of Washington DC's water supply coming from the Potomac River. More details and history of the WMA water supply system have been offered by Stagge and Moglen (2014) and Sheer and Flynn (1983).
Optimization of the WMA water supply system has its origins in the initial water-allocation studies (Palmer et al. 1979(Palmer et al. , 1982, which concluded that demand could be met through coordinated operation of the existing Patuxent and Occoquan Reservoirs, along with the Jennings Randolph and a then-proposed reservoir, which would eventually become the Little Seneca Reservoir. The system has been stressed several times, with water supply releases made on three occasions, in 1999, 2002, and 2010. Following the 1999 drought event, specific triggers were added to the management plan that guaranteed all regions (MD, VA, and DC) would enact water-use restrictions automatically and simultaneously to prevent jurisdictional disagreements. In an optimization study of the region, Stagge and Moglen (2014) concluded that these triggers were unnecessarily conservative, never engaging during simulations of the historical drought of record, and that accepting infrequent use restrictions would greatly decrease the system's vulnerability. Stagge and Moglen (2014) considered other water management rules, concluding that improvements to reservoir storage and environmental flowby could be achieved by modifying rules that shift demand from the Potomac River to the off-line reservoirs. Rules controlling the relative releases from the Jennings Randolph and Little Seneca Reservoirs were found to be relatively well optimized, although a slightly stronger reliance on releases from the Little Seneca improved overall storage and downstream flow targets.
Projections of climate change effects in the Potomac River watershed and mid-Atlantic United States predict moderate increases in mean annual temperature, precipitation, and streamflow over the next century (Najjar et al. 2009;Pyke et al. 2008;Hayhoe et al. 2008). An evaluation of the four best-performing general circulation models (GCMs) in the Chesapeake Bay watershed suggested an increase in mean annual temperature of 3.9 AE 1.1°C and an increase in precipitation of 9 AE 12% by the end of the century under the A2 Scenario (Najjar et al. 2009). This continues the historical trend of precipitation increases throughout the northeast United States during the twentieth century (Groisman et al. 2001(Groisman et al. , 2004. Despite projected increases in mean annual precipitation and flow for the mid-Atlantic, variation in the seasonality and distribution of precipitation and runoff is potentially more important for water resources management. Storm events are projected to become both more severe and intermittent, with precipitation intensity expected to increase by one standard deviation, concurrent with an increase in dry days and heatwaves (Meehl and Tebaldi 2004;Tebaldi et al. 2006).
These projections suggest a moderate increase in mean flows, but with greater likelihood of flooding resulting from storm intensity, and drought attributable to prolonged dry periods. Seasonality is also expected to shift, with the greatest increase in precipitation occurring during the winter and spring (Najjar et al. 2009). Similar seasonal trends were found by Mccabe and Ayers (1989), Moore et al. (1997), and Hayhoe et al. (2007). This was further supported by detailed simulations of flow in the Potomac River that projected a slight increase (1-7%) in mean annual flow by 2070-2099, with the increase occurring during the winter and early-spring peak season (Stagge and Moglen 2013). At the same time, summer flows are projected to decrease, caused by a decrease in runoff from large, sustained storm events, and the date of the minimum flow is expected to shift earlier by 2-5 days (Stagge and Moglen 2013).
In addition to climate change, demand increases and loss of storage due to sedimentation will further stress the system. The population of the WMA was predicted to increase by approximately 1 million people (25%) between 2010 and 2040, which corresponds to a projected water demand increase of 23% (MWCOG 2009). According to the most recent Census estimates (U.S. Census Bureau 2016), the region's population has already increased by 460,000 during the first 5 years of this period (2010)(2011)(2012)(2013)(2014)(2015). Adding to this potential system stress, reservoirs in the WMA water supply system are projected to lose 7-15% of their usable storage volume due to sedimentation in the 30 years between 2010 and 2040 (Ahmed et al. 2010).
This study has two primary objectives: (1) estimate future water supply vulnerability in the Potomac River and WMA, and (2) optimize water system rules based on future conditions and thereby provide adaptation strategies. The WMA represents an interesting challenge for this approach, given its tranboundary jurisdictional constraints and uncertainty because of the lag between reservoir releases and water delivery. Future conditions are simulated using the best available projections of demand change and reservoir sedimentation, whereas climate change effects are based on stochastically generated flows (Stagge and Moglen 2013) driven by Coupled Model Intercomparison Project Phase 3 (CMIP3) projections (Meehl et al. 2007). Adaptation strategies are derived by considering several conflicting objectives using start-of-the-art multiobjective evolutionary algorithm optimization. The advantage of this approach is a greater flexibility in objectives and system models that still allows decision makers to easily compare alternatives by metrics that are used in practice. The resulting strategies show how current levels of service in the WMA could be maintained in the future using only better management, avoiding the need for physical modification to the system. This demonstrates an approach merging climate projections and optimization that could be replicated in other water systems to develop adaptation strategies.

Methods
This study extends prior research on optimal water management on the Potomac River under current conditions Stagge and Moglen (2014) to instead test the vulnerability of the WMA water supply system to projected future climate, demand, and storage changes and then address the critical topic of adaptation to these future conditions. Future vulnerability was tested by comparing system performance using current conditions to three future climate periods (2010-2039, 2040-2069, and 2070-2099) and projections of demand and reservoir sedimentation at 5-year intervals from 2010 to 2040. Vulnerability was estimated for each of these scenarios separately and together, and performance was quantified using six objective functions considered in previous studies of the system. Adaptation strategies were determined by optimizing system rules using a multiobjective evolutionary algorithm approach and highlighting how optimal rules might mitigate vulnerabilities identified in the first part of the study. This study uses the water supply model developed and described in detail by Stagge and Moglen (2014). Hydraulic routing and reservoir operations were simulated using OASIS, which is a water management simulation and decision model that uses a node-arc architecture to model reservoirs, reaches, inputs, and withdrawals. Operating rules are expressed as goals or constraints and solved via linear programming using a daily time step, mimicking the imperfect foresight of daily operational decision making.
The OASIS model was developed in conjunction with the Interstate Commission on the Potomac River Basin (ICPRB) and water suppliers to ensure that all data, operating rules, and assumptions were accurate. Reservoir details, including stage-storage curves, sedimentation rates, and existing operational rule curves, were provided by the ICPRB, along with the current Potomac channelrouting and travel-time estimates. Daily demand among the three major WMA water suppliers was simulated using a set of multivariate regression equations, incorporating an autoregressive movingaverage (ARMA) error term, provided by Ahmed et al. (2010). Municipal water needs of the WMA are managed by three major suppliers: • Washington Suburban Sanitary Commission (WSSC), which serves the Maryland suburbs; • Fairfax Water, which serves Fairfax County and other northern Virginia suburbs; and • Washington Aqueduct, which provides water to the District of Columbia. The current water supply system ( Fig. 1) is the result of several design iterations and collaboration among the numerous levels of government, water suppliers, and citizen groups. The ICPRB's section for Cooperative Water Supply Operations on the Potomac (CO-OP), is responsible for coordinating water resources across these suppliers and stakeholders during times of low flow. Details of the system have been provided by Stagge and Moglen (2014) and Ahmed et al. (2010). This system relies predominantly (approximately 78% annually, Ahmed et al. 2010) on flow from the Potomac River to satisfy water demands, with the remainder of water provided by two off-line reservoirs: the Patuxent Reservoir system operated by WSSC and the Occoquan Reservoir operated by Fairfax Water (Table 1). Flow along the Potomac is augmented by two reservoirs. The Jennings Randolph Reservoir is the larger of the two (109 × 10 6 m 3 ), but is located approximately 9-10 days hydrologic travel time upstream of the WMA intakes (Table 1).
The Little Seneca Reservoir is located only 1 day upstream of the MWA intakes, but has significantly smaller usable storage and a smaller watershed area. These two reservoirs are, therefore, operated in concert, with the Jennings Randolph providing primary releases and the Little Seneca used to fine-tune flows immediately upstream of the intakes. The Savage Reservoir, located 8 km downstream from the Jennings Randolph Reservoir, is operated by the U.S. Army Corps of Engineers (USACE) in conjunction with the Upper Potomac River Commission (UPRC) to satisfy local North Branch low-flow requirements and supply water to the nearby town of Westernport, Maryland. It was not considered for optimization because it operates independently; however, the Savage Reservoir does make water supply releases during severe droughts according to a matching relationship with Jennings Randolph releases and therefore is also included in the model. This system layout possesses considerable uncertainty because release decisions must be made in advance of accurate weather forecasts yet still allow the main stem of the Potomac River to remain relatively uncontrolled.

Climate Change Flow Simulation
The effect of climate change was simulated by stochastically generating daily climate-adjusted streamflow and precipitation time series using the method described by Stagge and Moglen (2013). Five GCM models (Table 2) from the CMIP3 experiment (Meehl et al. 2007) were used to generate flows for three special report emissions scenarios (SRES A2, A1b, and B1). Projections of GCM-scale climate variables were related to discrete monthly climate states identified from the historical record for the study region. The Markov-chain transition probabilities between these climate states were then adjusted based on GCM climate projections. The parameters of a daily streamflow model, similar to those developed by Aksoy (2003) and Szilagyi et al. (2006), were defined by the monthly climate state and ultimately used to generate climate-adjusted daily streamflow. Daily flow was modeled using a two-state (increasing/decreasing) Markov chain, with rising limb increments randomly sampled from a Weibull distribution and the falling limb modeled as an exponential recession. This model was demonstrated to accurately reproduce historical streamflow statistics at the daily, monthly, and annual time steps in the Potomac River (Stagge and Moglen 2013) and to produce climate-adjusted streamflows that match the general findings of classical climate downscaling studies (Najjar et al. 2009;Milly et al. 2005;Hayhoe et al. 2007). Daily streamflow was generated for USGS Stream Gauge 01646500, located on the Potomac River near the Little Falls pumping station in Washington, DC and spatially disaggregated to daily streamflow and precipitation values at the necessary upstream sites using the Method of Fragments (Srikanthan and McMahon 1982;Porter and Pink 1991), in keeping with the approach of Stagge and Moglen (2014). Flows were bias-corrected using quantilequantile mapping to remove residual model bias, particularly at the upstream sites.

Demand and Sedimentation Projections
Demand projections (Table 3) were based on the most recent population and demand projections for the WMA (Ahmed et al. 2010). This projection evaluates demand change through the year 2040, modeling beyond the 20-year forecast legally mandated to be performed once every 5 years. These predictions are based on recent water-use information provided by the WMA water suppliers and demographic projections from the most recent Metropolitan Washington Council of Governments (MWCOG) Round 7.2 Cooperative Forecast (MWCOG 2009). Demand change beyond year 2040 is not considered in this study because water demand forecasts tend to become unreliable beyond the 30-year horizon in this region (Ahmed et al. 2010), given the added uncertainty of population change and innovations in water efficiency.
Sedimentation rates (Table 4) were based on historical trend analysis (Ahmed et al. 2010) using the Kendall-Theil Robust Line (Sen 1968). This nonparametric method is a popular alternative to linear regression and is more robust to outliers. The rate of sedimentation was assumed to remain constant for all future time steps, but was only projected until 2040 to match demand changes. This limit on the time horizon was meant to account for uncertainty in sediment capture methods or land-cover change.

Optimization of Operating Rules
Optimization of system operating rules was carried out in a manner similar to that used by Stagge and Moglen (2014) Zitzler andThiele (1998) andFleischer (2003), are invariant to objective scaling, tend to converge on the Pareto set, and assign a greater weight to regions with unique points or high curvature in the objective space. Optimization was carried out using the EMOA R package (Mersmann 2011) with simulated binary crossover (SBX) and polynomial mutation. This optimization scheme has proven efficient and effective relative to other multiobjective evolutionary algorithms in benchmark studies (Beume et al. 2007).
Within the range of available water resources optimization techniques, evolutionary, or genetic, algorithm solvers have proven successful because of their robustness and flexibility (Chen 2003;Momtahen and Dariane 2007;Oliveira and Loucks 1997;Wardlaw and Sharif 1999). Evolutionary algorithms are capable of searching large and complex decision spaces and evaluating nonlinear and nonconvex objective functions. Multiobjective evolutionary algorithm optimization solves for a set of compromise solutions, termed the Pareto optimal front, which represent optimal solutions that cannot be improved without affecting the other objectives.
Six objective functions were developed in conjunction with water suppliers and the ICPRB and designed to cover the range of potential benefits within the Potomac River system. Target volumes and flows were often based on legal agreements, such as the Low Flow Allocation Agreement (U.S. Army Corps of Engineers 1982). Because the functional limit of current multiobjective evolutionary algorithms has been shown to be approximately 10 objectives (Reed et al. 2013), this optimization model uses six objectives. Each objective is followed by the units of that objective in parentheses: 1. Shortage, which minimizes delivery shortages to the water suppliers (volume); 2. Storage, which minimizes low storage volumes in any of the reservoirs (volume); 3. Flowby, which minimizes days when flow in the Potomac does not exceed low-flow requirements (days of violation); 4. Rec Season, which minimizes days during the recreation season that Jennings Randolph levels fall below recreation facilities (days of violation);  5. Whitewater, which minimizes days when whitewater releases cannot be made because of low storage volume (days of violation); and 6. Env Flows, which minimizes days when flow in the Potomac falls below recommended environmental levels for three consecutive days (days of violation). These objectives are presented as a constrained multiobjective optimization problem, identical to that posed by Stagge and Moglen (2014)   During this period, water managers strive to maintain water levels in the Jennings Randolph Reservoir, represented as Elev JR , above three recreation access points. These points, termed E Beach , E WV , and E MD , are 443, 440, and 433 m, respectively. Z WW [Eq. (1f)] = ratio of days when whitewater releases, Q WW , cannot be made because of low storage volume. Whitewater releases are set to occur on the 15th and 30th of April and May, whose set is represented as T WW . Z Env Flows [Eq. (1g)] uses a measure to summarize water supply activity's effect on the ecological health of the Potomac River.
Although the legal flowby requirement below Little Falls is set at 757 × 10 3 m 3 =day, the Potomac Basin Large River Environmental Flow Needs study stated that there "is strong concern that a continuous, multi-day period of flows at or very close to 379 × 10 3 m 3 =day would be injurious to the biota" (Cummins et al. 2010). This function sums the number of occurrences when flow below Little Falls, Q LF , remains below 757 × 10 3 m 3 =day for three or more consecutive days. Five operating-rule modifications were considered based on recommendations by water suppliers and stakeholders. These rule modifications span a range of typical water management and conservation approaches and are identical to those considered by Stagge and Moglen (2014): (1) the buffer equation that shifts load between the upstream (Jennings Randolph) and downstream (Little Seneca) mainstem Potomac reservoirs; (2) load shifting, which shifts load from the Potomac to the off-line reservoirs; (3) metropolitan demand restrictions; and seasonal reservoir-release rule curves for the (4) Jennings Randolph and (5) Patuxent Reservoirs. Each candidate rule was optimized separately to determine its potential adaptation effect. Adaptation rules were generated using both the historical record and the Commonwealth Scientific and Industrial Research Organisation (CSIRO) (Gordon et al. 2002(Gordon et al. ) A2 scenario (2070(Gordon et al. -2099, both subject to year 2040 levels of demand and sedimentation. The CSIRO output was chosen as representative of SRES A2 conditions at the end of the next century, and the A2 Scenario was chosen as the most extreme case. In verification tests, the CSIRO model consistently produced good statistical agreement with the historical record across daily, monthly, and annual time steps.

Projected Changes to WMA Reliability
Three major processes are projected to affect the reliability of the WMA water supply system over the next century. These are demand change, reservoir sedimentation, and climate change. To identify the relative impact of these processes on the system, the system was simulated while adjusting to each parameter in isolation.

Vulnerability Caused by Demand Change
Demand forecasts predict a population increase of approximately 1 million (25%) between 2010 and 2040, which corresponds to a projected water demand increase of 430 m 3 =d (23%; Table 3) (MWCOG 2009). The greatest increase in population, and therefore water demand, is projected to occur within Fairfax Water's service area of northern Virginia. Demand increase for Fairfax Water is projected to increase by 31% between 2010 and 2040, whereas the WSSC and Washington Aqueduct service areas are expected to see increased demand of 19 and 18%, respectively. The City of Rockville, Maryland, which maintains a separate water supply, is projected to have a relatively large increase in demand by percent (31%), but this remains a small portion of the total WMA water supply because of Rockville's small service area.
This projected increase in demand will produce a consistent increase in storage penalty failures, Z Stor , and recreation season failures, Z RecSeason (Fig. 2). However, impacts are different, with sedimentation strongly affecting available storage [ Fig. 2(a)], and increased demand strongly affecting recreation season storage [ Fig. 2(b)]. By 2040, this increase in demand alone will result in an additional loss of approximately 0.5 days=year with access to the beach (2.0% increase) and 0.9 days=year with access to the West Virginia boat ramp (58.3% increase). Although this loss of recreation time may not appear large, a 58.3% increase in the more severe West Virginia boat-ramp failures suggests that demand will drive a loss of recreation revenue. Additionally, recreation failures tend to occur in extended groups, rather than a single instance. In this way, the additional failures may have a considerable effect on individual recreation seasons. Although increased demand does not dramatically affect WMA storage across all reservoirs [ Fig. 2(a)], by year 2030, it begins to adversely affect storage in the Little Seneca Reservoir, shown as an increased deviation between sedimentation-only scenarios and combined sedimentation and demand.

Vulnerability Caused by Sedimentation
Usable reservoir storage volume is expected to decrease because of the deposition of sediment carried by reservoir inflows over time. Reservoirs in the WMA water supply system are projected to lose 7-15% of their usable storage volume because of sedimentation in the 30 years between 2010 and 2040. Based on the most recent survey, the sedimentation rate in the Jennings Randolph Reservoir is particularly high relative to the other reservoirs (Table 4), and much greater than the original design sedimentation rate of 25 m 3 =year (Burns and MacArthur 1996). By year 2040, the storage-capacity loss in the Jennings Randolph Reservoir is projected to be 25% of the original storage volume (14.1% between 2010 and 2040). Despite these predictions of storage loss, sedimentation rates tend to change with time, as the sediment contribution of upstream watersheds change. Increased development tends to increase sediment load per area (Allmendinger et al. 2007), although this effect may be mitigated by improvements in nonpoint-source runoff treatment. The Jennings Randolph watershed, historically home to coal mining, has seen a decrease in this industry and has been subject to increased oversight with respect to nonpoint-source runoff.
Reservoir sedimentation is expected to increase the frequency and severity of reservoir storage failures, defined as usable storage less than 40% by Z Stor (Fig. 2). This increase is attributable primarily to storage failures in the Patuxent and Savage Reservoirs. Interestingly, the Jennings Randolph and Little Seneca water supply reservoirs do not develop storage failures until the year 2040 sedimentation level. This suggests that there may be opportunities for improving Z Stor when storage becomes lost to sedimentation through changes in how load is allocated among the reservoirs. Because Z RecSeason is strongly tied to storage in the Jennings Randolph, it is not surprising that Z Rec Season is relatively unaffected by sedimentation losses (Fig. 2). Furthermore, sedimentation has little impact on flow measures Z Flowby and Z Env Flows .

Vulnerability Caused by Climate Change
Output from five GCM simulations (Table 2) was used to generate streamflow and precipitation throughout the Potomac watershed at 30-year intervals (2010-2039, 2040-2069, and 2070-2099). These simulations predict a slight increase (1-7%) in mean annual flow over the next century, with increases during the winter and early spring, followed by decreased flow during summer (Stagge and Moglen 2013;Najjar et al. 2009;Hayhoe et al. 2007). Projections also show that summer flows will be characterized by longer periods of low flow (Tebaldi et al. 2006), with shorter but more intense storm events and an earlier occurrence of the annual minimum flow. As expected, the highest emission scenario, SRES A2, produced the most severe shifts in streamflow, whereas the low-emission scenario, SRES B1, produced a more modest change.
The effect of climate change alone on water supply reliability in the WMA region is shown graphically in Fig. 3. Climate change simulations project an increase (worsening) for nearly all objective functions over the next century. Results presented in Fig. 3 account for model bias by using quantile-quantile bias correction and always comparing projections against current conditions simulated using the same GCM. Interestingly, the greatest change for most objective functions occurs during the first part of the upcoming century (2010-2039), despite streamflow trends continuing consistently until 2099 (Stagge and Moglen 2013).
When examined in greater detail, the climate change scenarios result in an increase in the frequency of Patuxent and Savage storage failures, although the severity of these failures actually tends to decrease throughout the century. This is partially because load is shifted to other reservoirs such as the Little Seneca and the Occoquan, which previously did not produce storage failures, but begin to once subjected to climate change streamflows. Although storage in the Jennings Randolph Reservoir is never low enough to be considered a storage failure, climate change conditions greatly decrease the number of days with access to the Jennings Randolph

Adaptation Strategies
As expected based on the vulnerability portion of this study, runs combining the climate projections of the 2070-2099 A2 emissions scenario with 2040 demand change and sedimentation was the most challenging scenario for the WMA system. The value of implementing adaptation strategies to this extreme case was determined by comparing system penalties (objective function values) using optimized rules to current rules (Table 5). These results show that adjustments to the buffer equation can produce the greatest improvement under future conditions for most objectives. Load shifting to reservoirs off the mainstem offers modest improvements, primarily to the flowby penalty, whereas modifying demand restricts produces the smallest impact. Modification of the Jennings Randolph (JR) rule curve is effective for addressing objectives related to recreation storage and Potomac low flows, and Patuxent rule-curve modifications decrease reservoir storage penalties. No system shortage failures were noted and were, therefore, not included in the discussion. This is because the existing operating rules prioritize satisfying daily demand at the expense of violating the other objectives.

Buffer Equation
Within the WMAwater supply operating rules, the buffer equation is designed to balance storage levels between the reservoirs on the main stem of the Potomac River, the upstream Jennings Randolph Reservoir, and downstream Little Seneca Reservoir. Reservoir releases are calculated based on estimated demand; however, the buffer equation adds a so-called buffer flow to Jennings Randolph releases to account for imbalance in percent usable storage between the Jennings Randolph water supply volume and downstream Little Seneca storage. The existing buffer equation is represented by a solid diagonal line in Fig. 4, in which a negative storage imbalance recommends a larger than necessary release from the Jennings Randolph to reduce load on the Little Seneca. The right side of these plots (positive imbalance) reduces Jennings Randolph releases under the assumption that the deficit will be satisfied through releases from the downstream Little Seneca Reservoir. Under the current policy, the slope of the buffer equation [Figs. 4(a and b)]  is linear for both of these situations, with a maximum buffer flow of 568 m 3 =d.
Modification of the buffer equation produced the largest improvement of the considered modifications for future conditions, reducing the frequency of missed flowby targets (Z Flowby ) and number of consecutive days with extreme low flows (Z EnvFlows ) ( Table 5). Buffer-equation adjustments were partially capable of mitigating the impact of climate change, reducing most penalties for the 2070-2099 scenario to levels simulated with only demand and sedimentation. However, no version of the buffer equation was capable of reducing systemwide penalties under climate change, demand increase, and sedimentation to current levels.
The buffer equation reduces Z Flowby and Z EnvFlow failures by increasing the buffer flow when usable Little Seneca storage (%) is lower than that of Jennings Randolph [ Fig. 4(a)]. Under these optimized rules, a much greater release is made from the Jennings Randolph Reservoir in this situation, which in turn reduces load on the Little Seneca Reservoir and acts as a pulse in the Potomac River to prevent extreme low flows downstream of Little Falls. Similar recommendations were made for current climate conditions (Stagge and Moglen 2014), and the shape of the optimal buffer equation does not change substantially with time between current conditions and the 2070-2099 projection.
Although the right side of the equation has little effect on Z Flowby , it is important for improving Z RecSeason [ Fig. 4(b)], particularly for the 2070-2099 projection. This extreme scenario produced the most stress on the Jennings Randolph storage, where recreation storage is measured. Therefore, it follows that a lower buffer equation on the right side would reduce Jennings Randolph releases when storage is low relative to other reservoirs, thereby protecting recreation storage.

Load Shifting
Whereas the buffer equation deals with balancing releases along the Potomac River, load shifting controls how demand is allocated to the offline reservoirs, the Patuxent and Occoquan. When predicted flow in the Potomac River is not sufficient to satisfy predicted demand, production at the Patuxent and Occoquant water-treatment plants is temporarily increased above typical production levels. Following this load-shifting event, production at the offline reservoirs is curtailed an equivalent amount in order to replenish storage. Load shifting occurs only when storage in the Jennings Randolph, Little Seneca, Occoquan, and Patuxent remains above trigger points, called load-shift storage indices.
Modification of the storage indices and load-shift equation has relatively little impact on the WMA system in simulations of future demand/sedimentation conditions and climate change (Table 5). Although changes to load shifting generally results in better performance than the current policy, this improvement cannot completely mitigate the effects of either climate change or demand and sedimentation change. No trends exist over time among the optimized load-shifting parameters, suggesting that the effectiveness of load shifting has been maximized and that no further improvements will be realized with time.
Adjustments to the load-shift equation were shown to be effective under current conditions because the Occoquan Reservoir had unused storage, which could be used to reduce load on the already stressed Patuxent Reservoir (Stagge and Moglen 2014). However, as future conditions further constrain and stress the WMA system, the additional Occoquan storage is not as readily available, as shown by increases in Occoquan storage penalties (storage < 40%). Increasing the load-shift storage indices was another method of decreasing load on the stressed Patuxent Reservoir under current climate conditions (Stagge and Moglen 2014). However, under future conditions, this puts undue strain on the Little Seneca Reservoir, suggesting that the benefits of this approach are already maximized.

Monthly Rule Curves
All reservoirs in the WMA water supply system operate, at least during a portion of the year, according to zone-based rule curves, except for Little Seneca, which maintains a full storage volume throughout the year. To determine adaptation potential, operating rule curves for the Jennings Randolph and Patuxent Reservoirs were evaluated using multiobjective optimization. The Jennings Randolph Reservoir was chosen for evaluation because it is the primary water supply reservoir on the Potomac River, whereas the Patuxent Reservoir was most vulnerable to storage failures. Jennings Randolph water quality storage is managed by the Baltimore District of the U.S. Army Corps of Engineers and uses three zonebased rule curves (high, medium, and low) to guide water quality releases during the non-recreation-season months (September-April). These releases are designed to approximate the natural contribution of the Potomac River's impounded North Branch while refilling the reservoir prior to the summer recreation season.
Modifications of the Jennings Randolph rule curves primarily improved objectives related to Jennings Randolph storage (Table 5) little effect on storage failures because these primarily occurred in other reservoirs or during the summer season when the seasonal rule curves are not in effect. The projected climate change shift toward higher flows during the winter and spring, followed by lower flows in the summer and early fall, was mirrored by the optimized Jennings Randolph Reservoir rule curves. The optimized curves increased trigger points between March and May, immediately prior to the recreation season, forcing the Jennings Randolph Reservoir to operate more conservatively, making smaller releases during this time. In this way, the increase in spring flows is used to increase the storage buffer prior to a summer flow regime characterized by more severe low flows. Modification of the Patuxent rule curve is designed to maintain adequate storage in the highly stressed Patuxent Reservoir while also providing additional water supply for the WSSC. Simulations suggest that the Patuxent Reservoir is vulnerable during future droughts, typically entering low storage (<40%) conditions before the remaining WMA reservoirs and thereby contributing to the Z Stor penalty. For future conditions, adjusting the Patuxent rule curves improves Z Stor by 6.1-6.4% (Table 5). The Patuxent Reservoir operates using two rule curves that control daily water-treatment withdrawals based on storage zone. The adaptation improvement is attributed to an increase of approximately 1,000-1,500 × 10 3 m 3 in both the upper and lower rule curves between the months of September and February. This modification allows the Patuxent Reservoir to refill more effectively if storage is low during the fall and winter by decreasing water-treatment rates and shifting load back to the Potomac River. Although this shift is similar in both the climate change simulation and the sediment and demand-change simulation, the optimal rule curves deviate in midsummer. Likely because of increased summer drought severity attributable to climate change, the optimized upper and lower Patuxent rule curves for this scenario tend to be approximately 300 × 3 m 3 higher through the months of July and August. This allows the Patuxent Reservoir to operate even more conservatively for the most extreme scenario.

Demand Restrictions
The Metropolitan Washington Council of Governments standardized the implementation of water-use restrictions by setting three demand-restriction levels: voluntary, mandatory, and emergency, each with a unique storage trigger (MWCOG 2000). As part of the MWCOG agreement, all regional governments agreed to abide by these triggers, declaring restrictions simultaneously. Voluntary restrictions are triggered when combined storage in the Jennings Randolph and Little Seneca Reservoirs falls below 60%. Trigger points for mandatory and emergency restrictions are set at 25 and 5% for Jennings Randolph or Little Seneca storage, respectively (Table 6). This is a simplification of the actual MWCOG demandrestriction rules, but matches actual operations very well.
In a review of the WMA under current conditions, Stagge and Moglen (2014) found that the existing MWCOG demandrestriction triggers would never be implemented during a repeat of the historical streamflow record with current demand levels. Because stress on the WMA water supply increases with time, the likelihood of demand restrictions increases, highlighting the importance of an effective demand-restriction policy. Under the existing MWCOG policy and 2040 demand and sedimentation levels but no climate change, the WMA service area would experience voluntary restrictions once every 26 years, on average. Simulations based on the CSIRO 2070-2099 A2 climate scenario with demand change and sedimentation increase this frequency to once every 20 years, with 75% of voluntary restriction years ultimately requiring mandatory demand restrictions.
Improvements because of demand restrictions are limited and primarily focus on Z Flowby and Z EnvFlows . With regard to storage, these changes particularly improve storage in the Patuxent and Occoquan Reservoirs. System performance is improved by increasing the voluntary trigger from 60% of Jennings Randolph and Little Seneca storage to 74-85% (Table 6). Operations also improved when the mandatory restriction trigger point was decreased from 25 to 17-25% for Jennings Randolph storage but increased from 25 to 24-59% for Little Seneca storage ( Table 6). The trigger point is higher for the Little Seneca because it is more vulnerable because of its small size and slow refill rate. Trigger points for emergency restrictions were also increased, although these were so infrequently used that there is significant uncertainty in the results. The benefits of these adaptation strategies are tempered by an increase in the frequency of demand restrictions, for example, doubling the frequency of voluntary restrictions from once every 20 years to once every 10 years.
Modifying the percent demand restrictions during the summer season (June-September) did not produce significant improvement in the objective functions. However, some improvements for Z Flowby and Z Env Flows were realized by increasing the percent demand restrictions outside of the summer period to resemble summer restrictions. Continuing the more severe restrictions outside the summer drought period allowed reservoirs to refill prior to the next summer, better handling multiyear droughts.

Discussion
This study uses evolutionary algorithms to optimize water management strategies. However, other alternatives exist and could be substituted into this framework to identify adaptation strategies. More traditional optimization techniques such as linear or nonlinear programming have the benefit of quick convergence to the global optima, but would require several simplifying assumptions with regard to constraints, objectives, and adaptation strategies (Labadie 2004). More recent heuristic optimization techniques could also be considered, such as particle-swarm optimization (Reddy and Nagesh Kumar 2007;Taormina and Chau 2015), fuzzy programming (Chen and Chang 2010), or simulated annealing (Li and Wei 2008). Similar to the evolutionary algorithm approach used here, these alternative optimization approaches add a great deal of flexibility, sacrificing the guarantee of finding global optima and requiring more processing time. More detailed comparisons of modern optimization techniques are available in several methodology overviews (Ahmad et al. 2014;Sahinidis 2004;Labadie 2004). From among these alternatives, evolutionary algorithms were chosen because they are one of the most common heuristic optimization techniques and are proven to be robust, flexible, and capable of searching large and complex decision spaces (Reed et al. 2013). Flexible optimization schemes are important in complex systems like the WMA because they can be directly linked to hydrologic models and can handle uncertainty caused by time lags in water delivery and complex objective functions.
The objectives in this study were selected in close collaboration with the water suppliers and were designed to closely match the goals of the system as codified in legal agreements. However, there would be a benefit to considering new and more complex objective functions to determine how the set of optimal solutions would change. For example, the environmental and low-flow objectives are based on quite simple legal requirements, but the objectives could be better targeted to ecological health by collaborating with ecologists and fisheries experts. Similarly, there may be some benefit to considering more complex economic drivers and objectives, using a framework similar to that of Harou et al. (2009).
This study used CMIP3 projections downscaled to daily streamflow using the method of Stagge and Moglen (2013) rather than more traditional approaches, such as statistical or dynamical downscaling. The benefit of the Stagge and Moglen (2013) approach is that it generates a suite of ensemble members to better test vulnerability over a wider range of feasible flows and does not require a full hydrologic model. As described by Stagge and Moglen (2013), the existing Potomac River model performed poorly for low flows, whereas the alternative approach better captured these. The CMIP3 set of GCM runs has been updated with CMIP5 output (Wuebbles et al. 2014). It would be helpful to consider CMIP5 output in the future, although the two experiments agree well with regard to precipitation and drought near the Potomac River (Wuebbles et al. 2014). The largest improvements have been for simulation of monsoon precipitation, which mainly affects more southern and western parts of the United States (Cook and Seager 2013).
for communication of program information (Braille, large print, audiotape, etc.) should contact USDA's TARGET Center at (202) 720-2600 (voice and TDD). To file a complaint of discrimination, write to USDA, Director, Office of Civil Rights, 1400 Independence Avenue, S.W., Washington, D.C. 20250-9410, or call (800) 795-3272 (voice) or (202) 720-6382 (TDD). USDA is an equal opportunity provider and employer. The authors would like to thank two anonymous reviewers for their constructive comments regarding this paper.