Response of electricity sector air pollution emissions to drought conditions in the western United States

Water is needed for hydroelectric generation and to cool thermoelectric power plants. This dependence on water makes electricity generation vulnerable to droughts. Furthermore, because power sector CO2 emissions amount to approximately one third of total US emissions, droughts could influence the inter-annual variability of state- and national-scale emissions. However, the magnitude of drought-induced changes in power sector emissions is not well understood, especially in the context of climate mitigation policies. Using multivariate linear regressions, we find that droughts are positively correlated to increases in electricity generation from natural gas in California, Idaho, Oregon, and Washington; and from coal in Colorado, Montana, Oregon, Utah, Washington, and Wyoming. Using a statistical model, we estimate that this shift in generation sources led to total increases in regional emissions of 100 Mt of CO2, 45 kt of SO2, and 57 kt of NOx from 2001 to 2015, most of which originated in California, Oregon, Washington, and Wyoming. The CO2 emissions induced by droughts in California, Idaho, Oregon, and Washington amounted to 7%–12% of the total CO2 emissions from their respective power sectors, and the yearly rates were 8%–15% of their respective 2030 yearly targets outlined in the Clean Power Plan (CPP). Although there is uncertainty surrounding the CPP, its targets provide appropriate reference points for climate mitigation goals for the power sector. Given the global importance of hydroelectric and thermoelectric power, our results represent a critical step in quantifying the impact of drought on pollutant emissions from the power sector—and thus on mitigation targets—in other regions of the world.


Introduction
Electricity generation requires water resources to drive turbines in hydroelectric dams and to cool thermoelectric power plants that are fueled by nuclear, coal or natural gas. This dependence on water makes the electricity sector vulnerable to droughts (van Vliet et al 2012, 2016b, Bartos and Chester 2015, Voisin et al 2016, Gleick 2017, Hardin et al 2017, Miara et al 2017, Eyer and Wichman 2018. Hydroelectric dams and thermoelectric power plants supply a large fraction of electricity in the western United States, including 23% and 62%, respectively, in 2015 (EIA 2017). The energy portfolio varies across the region (EIA 2017) (see figure 1), but generally relies on a mix of hydroelectric and thermoelectric power for 'baseload' generation (Bartos and Chester 2015). When demand increases, additional coal, natural gas, and petroleumfueled power plants are dispatched to supply 'peaking' generation (Bartos andChester 2015, Miara et al 2017). The order in which power plants are dispatched in the US follows their variable operating costs, with lower-cost plants generally being dispatched first (Sioshansi 2008). Due to their lower operating costs, solar and wind are generally dispatched together with baseload power plants. While the relative fractions of hydropower to thermoelectric power plants vary from region to region, the dispatch of electricity generation from baseload to peaking power plants in order of variable operating cost is common across the world (IEA 2016a, IEA 2017).
By reducing streamflow, droughts can decrease instate baseload electricity generation from hydroelectric and thermoelectric sources (Bartos and Chester 2015), thus requiring dispatch of peaking generation plants ( Bartos andChester 2015, Miara et al 2017) to prevent brownouts or blackouts. Droughts may also be accompanied by heat waves (e.g. Mazdiyasni and AghaKouchak 2015), requiring dispatch of peaking power plants to meet electricity demand for air conditioning. High air temperatures may lead to warmer streamflow, which can reduce generation capacity from thermoelectric plants via regulatory constraints on the use of streamflow for cooling (van Vliet et al 2012(van Vliet et al , 2016a(van Vliet et al , 2016b. Depending on the region, cold temperatures during winter may also increase electricity demand for heating. The increased use of fossil fuel power plants for peak generation induced by droughts may last from months to years, which can lead to significant increases in pollutant emissions from the electricity sector (Gleick 2017, Hardin et al 2017, Eyer and Wichman 2018. In addition, decreased in-state electricity generation due to droughts may increase the need to import electricity from neighboring states (van Vliet et al 2013, Voisin et al 2016), potentially causing remote increases in pollution.
Past studies have coupled models of surface hydrology, streamflow temperature, and power generation to explore how changes in streamflow and river temperatures affect electricity generation from hydroelectric and thermoelectric plants in current and future climates (van Vliet et al 2012, 2016b, Bartos and Chester 2015, Voisin et al 2016, Miara et al 2017. Other studies used a generation expansion planning model to explore the tradeoffs between water withdrawals, air quality, and electricity generation in the context of planning future generation capacity , or applied a plant-scale power generation model over a region to explore the optimal changes in electricity dispatch during droughts (Pacsi et al 2013). These modeling approaches tend to focus on either the hydrologic risks faced by individual power plants, or the optimal regional operations of electricity systems in response to water resources constraints.
Recent plant-scale econometric analyses found a positive relationship between water scarcity and emissions from the US electricity sector (Eyer and Wichman 2018), and for the recent California drought (Gleick 2017, Hardin et al 2017. Eyer and Wichman (2018) found that in the Western Interconnection,  (2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015) of in-state net-generation from hydropower, natural gas, coal, nuclear, petroleum, and solar and wind (combined); and average yearly deficits and surplus of in-state electricity generation compared to total electricity sales. Error bars represent standard deviation of the linearly detrended time series of annual totals (n=15), and dots show the underlying yearly time series for each source (before detrending). Note that California and Washington have a different y-axis scale. power plants that use once-through cooling and cooling ponds show the strongest relationships with water scarcity, as do power plants that use surface water and municipal water. However, it is still unclear what fraction of historical electricity sector emissions can be attributed to droughts, whether those emissions pose challenges to meeting climate and air quality mitigation targets, and how drought-induced emissions may be affected by future climate change.
Given these questions, we conduct a rigorous statistical analysis of the historical sensitivity of electricity sector emissions to drought. Our approach uses state-wide data on electricity generation and pollutant emissions, combined with a hydrologically-based representation of past droughts. This rigorous characterization of hydrological droughts is critical for adequately quantifying the impacts of climate variability on water availability for electricity generation.
The response of electricity sector emissions to drought is influenced by at least three factors: (1) the frequency, duration, and intensity of droughts; (2) the importance of water-dependent electricity sources in the total electricity portfolio; and (3) the mix of energy sources that replaces the hampered water-dependent power generation. To examine the interaction of these factors, we focus on the western US (figures 1-4), which provides an ideal test case for hydropower regions globally. First, the region relies heavily on hydropower and water- . Second, the electricity mix varies widely across states (figure 1), providing critical variation through which to examine the response to droughts. Third, the states are connected by the Western Interconnection sub-grid, which captures most of these states' electricity trade. Fourth, sub-regional data on electricity generation (EIA 2017) and pollutant emissions (EPA 2017) are available. And fifth, several important droughts have affected the western US in recent years, and it is expected that the region will experience increasing likelihood of droughts due to climate change, in part due to reduced water availability during the spring and summer associated with decreased snowpack ( . This regional drought prevalence provides multiple events through which to measure the response of the electricity sector. We use a statistical approach to quantify the impacts of droughts directly from observations (in this case from data on electricity generation and power and for the total for the electricity sector in each state (EIA estimates). Discrepancies between the two are likely due to emissionaccounting differences between the datasets. Yearly SO 2 and NO x emissions are only shown from natural gas and coal-fueled power plants that report to the EPA Clean Air Markets. Error bars represent standard deviation of the linearly detrended time series of annual totals (n=15), and dots show the underlying yearly time series (before detrending). sector emissions). Our empirical analysis adds to the growing climate econometrics literature, which studies the impacts of short-term climate shocks and long-term climate changes on human activities. Physically-based models often require numerous assumptions about factors such as water resources allocation, power plant operations and power plant efficiencies, which may introduce important biases. Statistical studies can thus aid the model development process by providing observational benchmarks that  can be used in model parameterization and/or validation.

Methods
We use multivariate linear regressions to estimate the impact of drought (characterized by negative runoff anomalies) on electricity generation and pollutant emissions from historical observations between 2001 and 2015. To quantify the relative importance of historical drought-induced emissions, we compare the drought-related CO 2 emissions that we estimate using a statistical model to actual state-level electricity sector totals and policy targets.

Data
We use monthly data from phase 2 of the North American Land Data Assimilation System (NLDAS-2), which provides data for the contiguous US at 0.125°r esolution, from 1979 to the present (Xia et al 2012a(Xia et al , 2012b. We use data from three NLDAS-2 land-surface models: VIC (Liang et al 1994), Noah (Chen et al 1996), and Mosaic (Koster and Suarez 1994). We define drought as periods of negative cumulative runoff anomalies, based on surface runoff and subsurface runoff (i.e. water transport through the soil that recharges streams and reservoirs from underground) (see section 2.2). While all three models show strong correlations with observed streamflow, VIC has been found to represent runoff more accurately than Noah (which overestimates runoff) and Mosaic (which underestimates runoff) (Xia et al 2012b). We also calculate population-weighted state averages of total degree day anomalies using NLDAS-2 daily mean air temperatures (see supplementary text 1, figures S1 and S2. These are available online at stacks.iop.org/ERL/ 13/124032/mmedia for details). We use monthly state-level electricity generation and sales data from the US Energy Information Agency (EIA; EIA 2017). We calculate electricity deficits and surplus in each state by subtracting the total net in-state electricity generation from total electricity retail sales. We obtain monthly power plant-level measurements of CO 2 , SO 2 , and NO x emissions data from the US Environmental Protection Agency (EPA 2017). We aggregate the power plant-level emissions to the state-level by adding the respective emissions from all power plants reporting in each state at each time step (figure 2). We also use yearly CO 2 emissions estimates for the whole electricity sector in each state, provided by the EIA (EIA 2018) (figure 2).

Characterizing drought
Runoff is the key hydrological variable that determines recharge of rivers and reservoirs, which are the main water sources for the electricity sector in the western US. Thus, calculating a drought index based on runoff data captures hydrological drought dynamics more accurately than standardized drought indices (e.g. Palmer Drought Severity Index) that may not capture the nonlinear rainfall-runoff relationships accurately due to their oversimplified representation of landsurface fluxes (Liang et al 1994, Chen et al 1996, Sheffield et al 2012, Trenberth et al 2014, Xia et al 2012a, 2012b. We add surface and subsurface runoff to calculate total runoff in each month for each NLDAS-2 grid-cell from 1979 to 2017. We create time series of monthly total runoff anomalies by subtracting the respective calendar-month mean from each individual monthly value in the original time series. This anomaly time series thereby quantifies the departure from the longterm mean that occurred in each month. To capture the longer time scales of hydrological droughts Wood 2011, van Loon 2015), we calculate a running sum of the monthly anomaly time series using accumulation windows of 3, 6, 9, and 12 months (figure S3). We find that the statistical models calibrated using the 12-month window yield lower errors compared to observations, so we use this window throughout the study (see supplementary text 2.4 and 3.2 for the full sensitivity analysis). Finally, we calculate state averages of the cumulative runoff anomalies for each month to match the spatial scale of the electricity generation data (figure S4).
We use a binary drought metric given by equation (1) to calculate a first approximation of the impacts of droughts on electricity generation: Here, d i,s is the value of the drought metric during month i in state s, q i s , is the 12-month cumulative runoff anomaly during month i in state s, and q s s  ( ) is the standard deviation of the 12-month cumulative runoff anomalies in state s calculated using the 1980-2017 time series.
We repeat the analysis using the full time series of 12-month cumulative runoff anomalies (see section 2.3 and supplementary text 2.2). We complete the analyses separately using data from VIC, Noah, and Mosaic to account for hydrological model uncertainty.

Statistical analysis
The electricity sector evolved across the western US from 2001 to 2015 due to a number of factors, including changes in natural gas prices relative to coal, in availability of solar and wind power, in demand due to population growth, in energy use efficiency, and in policies to reduce greenhouse gas emissions. These trends are evident in the time series of electricity generation and power sector emissions (figures 1 and 2). Because our goal is to quantify the relationship between year-to-year climate variability and power sector variables, we remove these long-term (often nonlinear trends) using piece-wise linear fits. We repeat the detrending exercise 17 times for each time series to test the sensitivity to the arbitrary choices made in the detrending algorithm (see details of the algorithm in supplementary text 2.1). We propagate this sensitivity analysis through the rest of the calculations, generating distributions of results. Thus, our analyses quantify the drought impacts on the variance of the residuals of the detrended time series.
We calculate multivariate linear regressions between the binary drought metric and the detrended time series of electricity generation (from hydropower, natural gas, coal, nuclear, and petroleum) and of surpluses and deficits of electricity for each state (see details in supplementary text 2.2). In addition, we repeat the multivariate linear regressions using time series of 12-month cumulative runoff anomalies. The general structure of the regressions is given by equation (2): where G s,t is the detrended time series of electricity generation from a given source in state s; Q s,t, is the time series of the binary drought metric or the 12month cumulative runoff anomalies; and T s,t is the time series of monthly anomalies in populationweighted total degree days in state s. λ t are controls for each month; β 0 is the regression intercept; β 1 and β 2 are the regression coefficients; and u t is the stochastic error term. Details of the calculation of absolute changes in generation from the regression coefficients can be found in the supplementary text 2.2. From these two sets of regressions, we identify groups of states where generation from natural gas is significantly (p<0.05) positively/negatively correlated with drought events (which we call 'NG+' and 'NG−', respectively). Similarly, we identify groups of states where generation from coal is significantly positively/negatively correlated with droughts (which we call 'C+' and 'C−', respectively). We detrend the time series of emissions for each state, and pool together the data of states within each of these four groups (NG+, NG−, C+, and C−) to extend the data available to study the impact that droughts of different magnitudes have on CO 2 , SO 2 , and NO x emissions. We sort the cumulative runoff anomalies into bins according to their magnitude (figure S4) and calculate multivariate linear regressions for the pooled emissions data similar to equation (2), with additional predictors to determine the effect of droughts of different magnitudes and with controls for the different states that are included (see supplementary Text 2.3 for details).

Calculating cumulative emissions
We use the results from the pooled, binned regressions described in section 2.3 to build a statistical model that predicts for each state how CO 2 , SO 2 , and NO x emissions changed given monthly values of cumulative runoff anomalies and total degree days anomalies from 2001 to 2015 (see supplementary text 3 for details). We repeat this procedure using regression results derived from VIC, Noah and Mosaic separately, and for the cascading sensitivities from the detrending algorithm, leading to 867 different estimates of drought-induced CO 2 , SO 2 , and NO x emissions per state per month from 2001 to 2015. To report the total CO 2 , SO 2 , and NO x emissions associated with negative cumulative runoff anomalies (i.e. hydrological droughts) in each state, we first add together the emissions from natural gas-and coal-based generation. We then add the respective monthly estimates for each state to obtain state-level total estimates of CO 2 , SO 2 , and NO x emissions throughout these 15 years (2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015), and calculate mean yearly rates by dividing these total estimates by 15.
This methodology captures the changes in each state's emissions caused both by replacing hampered in-state generation capacity and by increasing in-state generation to export electricity to neighboring states that are also affected by a drought (without differentiating between the two). Thus, these relationships apply for the conditions present between 2001 and 2015, including infrastructure, policies, electricity trade patterns, fuel prices, and spatial correlation of droughts.

Results
For each state, we calculate the distribution of average drought response of electricity generation from hydropower, natural gas, coal, petroleum, and nuclear over the 2001-2015 period (figures 3 and S5). We measure the performance of the statistical relationships by calculating the normalized root mean square errors (supplementary text 2.5) shown in figure S6.
The Western Interconnection also enables electricity to be traded within the region. The in-state generation deficit exhibits positive correlations in Idaho (+4%), suggesting increasing need for imports. Conversely, the in-state generation surplus exhibits negative correlations in Washington (−42%), Utah (−11%), and Oregon (−6%), suggesting a reduced ability to export; and positive correlations in Wyoming (+26%) and Montana (+17%). The pattern of drought response suggests that generation from coal increases in Wyoming and Montana in order to export electricity to surrounding states that experience drought-induced declines in generation.
We calculate multivariate linear regressions to estimate CO 2 , SO 2 , and NO x emissions linked to droughtinduced changes in generation from natural gas and coal in states that show similar responses (NG+, NG−, C+, C− figures S7-S11). We find that the largest increases in CO 2 emissions occur in California (median estimate of 51.3 Mt, total between 2001-2015, Washington (21.8 Mt), Oregon (13.5 Mt) and Wyoming (6.9 Mt) (figure 4 and table 1). Emissions in Washington amounted to 12.4% of the total power sector CO 2 emissions from the state, compared with 11.0% for Oregon, 7.4% for California, and 1.1% for Wyoming (which has relatively high state-level total power sector emissions; figure 2). On the other hand, the power sector in Idaho emits relatively little CO 2 , so although estimated drought-induced emissions for Idaho are only 1.0 Mt, these amount to 9.0% of its total power sector emissions. While the total emissions are key for climate change mitigation globally, the relative amounts for each state are important when developing state-specific emissions-reduction targets.
Arizona shows a negative relationship between coal-fueled generation and droughts, and a positive relationship between natural gas-fueled generation and droughts (figures 3, S5, and S7). The combination of these two effects results in 'negative' CO 2 emissions in our statistical model, but no changes in SO 2 or NO x emissions. However, it should be noted that there were few negative 12-month cumulative runoff anomalies in Arizona between 2001 and 2015 ( figure S4).
The spatial distribution of changes generally resembles the underlying electricity mix (figure 1), with states that either rely heavily on hydropower (e.g. Washington, California, and Oregon) or coal-fueled generation (e.g. Wyoming, Utah, and Colorado) producing higher emissions in response to drought. The size of California's electricity sector represents a unique case. For example, although hydropower supplies a smaller fraction of electricity in California than in Oregon (figure 1), and although California's hampered hydropower is replaced by a less carbon-intensive alternative (natural gas) than Oregon's (coal and natural gas) (figure 3), California's large total electricity demand (figure 1) means that California experiences greater increases in total CO 2 emissions during droughts (table 1).
We estimate the total 2001-2015 drought-induced emissions over the western US to be 100.1 Mt of CO 2 , 44.8 kt of SO 2 , and 56.9 kt of NO x . The annual-mean increase in regional CO 2 (6.7 Mt of CO 2 /year) is equivalent to the emissions of ∼1.45 million vehicles/ year (assuming 4.6 metric tons of emissions per vehicle/year, EPA 2018), or ∼5.7% of the automobiles Table 1. Historical emissions from electricity generation caused by droughts in each state. Median estimates (n=867) of total CO 2 , SO 2 , and NO x emissions from natural gas and coal-fueled power plants induced by droughts in each state between 2001 and 2015. Right columns for each pollutant represent the values as a percentage of the total historical emissions from the power sector (2001-2015) (figure 2). Median regional values are calculated from the distributions of regional totals (not the sum of the medians for each state). CO 2 percentages are relative to totals from the electricity sector. SO 2 and NO x percentages are relative to total emissions from natural gas and coal-fueled power plants that report to the EPA Clean Air Markets. The full distributions are shown in figure S12. Tons were rounded to the nearest ten.  figure S12 for the distributions of emissions per state, figure S13 for details on the California drought, and figure S14 for the sensitivity of regional emissions to the window of the cumulative runoff anomalies.) These drought-induced emissions are large enough to pose challenges to western states' progress towards their CO 2 emissions targets. Table 2 shows annual-mean (2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015), state-level drought-induced CO 2 emissions as percentages of (1) the states' total emissions from the electricity sector between 2001-2015, (2) targets of annual CO 2 rates established by the Clean Power Plan (CPP) (EPA 2015), and (3) state-specific targets of yearly CO 2 rates (supplementary table 1). While there is uncertainty regarding the implementation of the CPP targets, they provide policy benchmarks for electricity sector emissions. We find that drought-induced emissions can account for a significant portion of electricity sector emissions and related targets. For example, mean annual drought-induced emissions between 2001 and 2015 represent 15.0% of the 2030 CPP target for Washington, 12.2% for Oregon, and 7.8% for California. Thus, our results highlight the importance of the carbon intensity of the electricity sources that are used to replace hampered hydroelectric and thermoelectric generation during droughts.

Discussion and conclusions
We derive robust statistical relationships between runoff anomalies and state-level power sector emissions of CO 2 , SO 2 , and NO x . We characterize drought using runoff because it is an appropriate hydrological variable to study drought impacts on the power sector at the state-level, and NLDAS-2 models capture its behavior with high accuracy (Xia et al 2012b). Our results suggest that, between 2001 and 2015, power sector emissions attributable to droughts reached ∼10% of the average total annual power sector emissions in California, Idaho, Oregon, and Washington. Our results are in general agreement with recent analyses using plant-level data (Eyer and Wichman 2018), which found that natural gas is often used to replace hampered generation during droughts. However, Eyer and Wichman (2018) did not find strong relationships between generation from coal and water scarcity in the Western Interconnection. They also did not find statistically significant increases in CO 2 and NO x in the Western Interconnection, except when pooling data from across the country. We believe that this difference arises from the structure of our regressions, which allows for flexibility in capturing the response across the highly heterogenous energy mix within the Western Interconnection. By carrying out the analysis over each state individually, we identify both positive and negative state-level correlations between coal generation and drought, suggesting that the impact of drought on coal generation varies widely depending on the state. Creating separate statistical models for groups of states that show similar behaviors allows us to capture the heterogenous drought response across the region.
Climate change could alter the pattern of drought frequency and severity in future decades, particularly during the spring and summer due to reduced snowpack (Seager et al 2007, Sheffield and Wood 2008, Rauscher et al 2008, Diffenbaugh and Ashfaq 2010, Seager and Vecchi 2010, 2015  their historical levels (absent other interventions such as investments in renewable technology and thermoelectric power plants with more efficient cooling systems and emissions controls). California, Oregon, and Washington have established aggressive emissions-reduction targets through legislation (e.g. California ARB 2017), so even if the CPP is not implemented, emissions from the power sectors in these states may generally decrease as they strive to meet their state-level targets. However, our results raise the importance of considering drought-induced emissions from the power sector when developing mitigation policies. In addition to the first-order priority of replacing baseload generation with low-emission sources, meeting state-level goals may also require replacing peakingplants with low-emission sources. In California, generation from renewables is increasing, but nuclear fission power plants (a low-emissions energy source) are being retired (e.g. Davis and Hausman 2016). Unless enough renewables are brought online to replace the retired nuclear power, California could see a relative increase in drought-induced emissions as more natural gas power plants are brought online more frequently. Alternatively, California may drive up emissions in surrounding states if it increasingly relies on electricity imports from states that lack strong climate mitigation policies.
A recent report by the US Department of Energy outlined a roadmap to increase hydropower capacity by 50% by 2050, mainly by upgrading existing facilities, adding power generation capability to non-powered dams, and increasing pumped-storage (DOE 2016). Following this expansion raises the importance of considering the vulnerability of electricity generation to climate variability. Moreover, the issue of droughtinduced emissions is of global importance. Hydropower accounts for 16% of the world's electricity (including as high as 56% in Latin America; IEA 2016a, 2017), with an expected increase in worldwide generation of 47%-70% by 2040 from 2016 levels (IEA 2016b). In addition, thermoelectric power plants are responsible for 70% of the world's electricity (IEA 2016a, 2016b), of which fossil fuel power plants account for 58% of all energy-related water withdrawals (IEA 2016b) and 25% of all greenhouse gas emissions globally (IPCC 2014). Our study reveals the potential significance of drought-related emissions from the power sector, and highlights their importance for achieving emissions targets. In addition, the prospect of future increases in droughts globally (e.g. Wood 2008, 2011, Diffenbaugh and Berg and Sheffield 2018) raises the question of whether global-scale drought-induced emissions could be large enough to significantly impact global climate forcing, and hence drive further changes in regional climate and associated drought risk.