How Does Welfare from Load Shifting Electricity Policy Vary with Market Prices? Evidence from Bulk Storage and Electricity Generation

Transcription

1 How Does Welfare from Load Shifting Electricity Policy Vary with Market Prices? Evidence from Bulk Storage and Electricity Generation J. Scott Holladay Jacob LaRiviere March 2017 Abstract We model the electricity market to demonstrate that changes in the price of natural gas can cause the market and non-market impacts of bulk electricity storage to move in opposite directions. We provide evidence consistent with the model using a series of reduced form tests on data from We then simulate installing bulk electricity storage on the US electric grid. We find that lower natural gas prices generally reduce the market gains and non-market costs of storage. However, direct non-market costs are still positive which means that there is no argument for subsidizing storage to mitigate pollution given the current mix of generating technologies; arguments in favor of bulk storage R&D subsidies ride on public good aspects of technology and dynamic investment incentives for intermittent renewables. JEL Classification: H44, L5, L9, Q4, Q5, Keywords: Electricity storage, natural gas, air pollution, energy policy Department of Economics, University of Tennessee; 507 Stokely Management Center 916 Volunteer Boulevard Knoxville, TN 37996; jhollad3@utk.edu Department of Economics, University of Tennessee and Microsoft, Redmond, WA; jlarivi1@utk.edu 1

2 1 Introduction Policy makers in the United States have introduced but not yet passed legislation to subsidize bulk electricity storage research and development and bulk storage installation subsidies. 1 There are two main reasons why bulk storage subsidies might be socially desirable. First, subsidizing bulk storage implicitly subsidizes investment in intermittent renewable capacity. 2 Bulk storage effectively turns an intermittent renewable into a dispatchable generator thereby addressing the intermittency concerns that impact widespread renewable deployment. In doing so, bulk storage weakly increases revenue for intermittent generators because they are no longer pure price takers that can only supply when the wind blows or the sun shines but are strategic suppliers in the wholesale market. Thus, one benefit of bulk storage is the increased profitability of renewables and therefore more renewable investment. Insofar as renewables offset air pollution, subsidizing bulk storage decreases the externalities associated with the air pollution from the electricity sector as well (Cullen (2013), Kaffine et al. (2013), and Novan (2015)). Second, because discharging stored electricity is most profitable during peak load hours when prices are the highest, bulk storage decreases the net load that existing power units must satisfy during times of peak demand. 3 Therefore, bulk storage can mitigate the need for rarely used peaker units that can lead to a more efficient use of capital in the electric power sector. In the context of renewable generators, the non-market impacts of charging and discharging bulk electricity storage capacity have two components. Carson and Novan (2013) describe how bulk electricity storage directly leads to increased profits for renewable generators. Thus, bulk storage has an indirect public gain from increased renewable investment. While there is a large public gain from bulk storage through more investment in zero emission renewables, Carson and Novan (2013) also shows bulk storage could have a negative 1 Most notably, the introduction of the STORAGE Act for debate to US Congress on May 20, 2009, which offered large subsidies for installed storage capacity (Kaplan (2009)and DOE (2013)). 2 See and The Role of Bulk Energy Storage for Facilitating Renewable Energy Expansion released in 2012 by Energy Economics Group (EEG) of Vienna University of Technology, Austria. 3 California s duck curve is a notable exception to this general pattern. 2

3 direct impact on emissions. If the marginal emitter during charging hours is dirtier than the marginal emitter during discharging hours then storage leads directly to increased emissions. The sign of the direct emissions impact of bulk storage therefore depends on the marginal fossil fuel emitter when electricity is stored and discharged. Thus, the merit dispatch order of fossil fuel generators determines the direct non-market impacts of bulk storage. In this paper we investigate how direct private gains and theoretically ambiguous direct emission effects from bulk storage vary with fuel input prices. Recent decreases in natural gas prices because of horizontal drilling and fracturing extraction techniques (i.e., fracking ) motivate this line of research. 4 We focus on two interrelated questions: 1) how have lower natural gas prices impacted bulk storage s profitability through arbitraging wholesale electricity prices across off-peak and on-peak hours; and 2) how lower natural gas prices have affected bulk storages direct impacts on emissions (e.g., direct non-market incentives to invest in renewable generation capacity)? Put another way, we focus on how price changes in natural gas affect the market and non-market benefits of bulk storage. The first question relates to the potential private gains from bulk storage for renewable generators, and therefore renewable investment decisions. We don t address the more challenging problem of building a dynamic model of renewable investment which quantifies the precise indirect emission decreases that subsidized bulk storage would provide. 5 In our theoretical model, we extend Carson and Novan (2013) to show that the relative market and non-market direct effects of input price fluctuations on bulk storage are a joint function of a) input prices and b) the distribution of relative power unit productivity for each type of fossil fuel generation. Heat rates (e.g., the energy content of fuel used to gen- 4 Fracking is the process of forcing hydraulic fluid down a well shaft under pressure, which breaks apart porous rock releasing small amounts of trapped gas. Fracking, along with improvements in horizontal drilling, in which a drill bit is sunk up to 2 kilometers and then turned at a near 90 degree angle, have led to large increases in natural gas production from previously inaccessible deposits. See Joskow (2013) for a description of the fracking technologies and their effect on natural gas prices. 5 While worthwhile, that focus requires measuring the correlation between marginal emissions and renewable generation profiles similar to Callaway et al. (2015) in addition to developing a dynamic investment model as in Gowrisankaran et al. (2016). Our research design thus focuses on the basic question of arbitrage and direct forgone emissions conditional on a renewable generator being built. 3

4 erate one KWh of electricity) and fuel input prices jointly determine the marginal cost of a power unit s electricity. 6 Because electricity is generally dispatched in order of marginal cost, these two factors dictate the marginal generating unit, marginal emissions and wholesale electricity prices for different load levels. 7 When natural gas prices fall, highly efficient natural gas power units can become competitive with inefficient coal units and displace some coal generation. As natural gas prices fall the model predicts that dispatch of coal generators become more sensitive to their heat rates. Finally, a parameterized version of the model provides closed form solutions for direct market and non-market impacts of bulk storage as a function of changing input prices. We subsequently implement two sets of empirical exercises. First, we confirm that fossil fuel generators responded to natural gas price changes as predicted by the model. We estimate the distributions of the productivity for coal and natural gas power units from the EPA s Continuous Emissions Monitoring System (CEMS) data. We then estimate fossil fuel plant capacity factors as a function of its heat rate for different price levels of natural gas. We find that low natural gas prices are associated with the capacity factors of coal units becoming more sensitive to their heat rates and natural gas units becoming less sensitive. Put another way, a coal unit s efficiency becomes more important for its dispatch decision as natural gas prices fall. This is both unsurprising and consistent with the predictions of the model. However, a number of other possible channels could generate such a result so we do not claim causality. Second, using the closed form solutions from the theoretical model we simulate the direct public and private impacts of installing a small amount of bulk storage at different locations throughout the United States as a function of natural gas prices. We use the EPA s CEMS data along with wholesale electricity price data collected from the Federal Energy Regulatory Commission (FERC) and PJM, a wholesale market, to simulate storage charging and discharging behavior. We then combine that data with results on marginal 6 Operation, maintenance, depreciation, and environmental costs also affect the marginal cost of operation but it is common in the dispatch model-based analysis to assume they are jointly second order. 7 Sioshansi et al. (2009) uses an engineering approach to estimate private returns to bulk storage in a single electricity market, highlighting the importance of the fuel mix of existing generators. 4

5 emissions over time from Holladay and LaRiviere (2016). 8 We find that almost everywhere in the United States the market and non-market benefits from using bulk storage with wind generation move in opposite directions as natural gas prices decrease. Our simulations show that the non-market direct impacts of storage become more favorable while the arbitrage opportunities from bulk storage become less favorable when natural gas prices declined. The reason is straightforward: with lower natural gas prices, natural gas generation now competes with coal for baseload generation at night. As a result, charging hours are relatively cleaner than when natural gas was more expensive. This result leads to more favorable direct emission impacts from bulk storage. However, the total direct benefits still lead to emission increases. Conversely, wholesale electricity prices, especially during peak demand hours, are much lower with low natural gas prices. 9 As a result, the market benefits from arbitraging are much lower with inexpensive natural gas. The results have important policy implications. First, we provide evidence that the decrease in natural gas prices has significantly decreased the direct non-market costs of bulk storage through much of the U.S. Second, any bulk storage subsidy needs to be larger when natural gas prices are low to motivate widespread adoption since arbitrage opportunities are lower. We thus provide evidence that evaluating the robustness of gains and costs from energy policies due to input price fluctuations is a possibly overlooked criterion for policy-makers. This overlooked criterion increasingly matters in the electricity sector as power units pay a greater share of electricity with natural gas whose prices are more volatile than the coal prices that they pay (Chu et al. (2017)). Finally, our findings highlight the uncertainty in benefits from policies that support specific technologies in the energy sector as opposed to pricing externalities directly. 8 Holladay and LaRiviere (2016) examines the impact of inexpensive natural gas on marginal emissions rates and the environmental benefits of renewable generation. It does not consider bulk storage or the electricity price impacts of changes in natural gas prices. 9 Due to the fall in wholesale electricity prices, the benefit of many investments in electricity generation, including nuclear has likely decreased as well (Davis (2012)). 5

6 2 Background Bulk electricity storage allows electricity to be generated at one time of day and sold during another. The private benefit of bulk storage is that it allows inexpensively generated electricity to be dispatched during peak demand hours when wholesale electricity prices are high. This both increases profits for off-peak electricity generators who store their generation and decreases peak wholesale electricity prices. 10 Because bulk storage shifts the net load, there is an effect on unregulated pollutants from the electricity sector. Therefore, any welfare calculation of the direct effects of bulk storage must account for increases or decreases in emissions from adding bulk storage to the grid. 11 There are multiple competing bulk storage technologies. Most units use pumped hydroelectric storage in which they pump water up hill and generate electricity when the water runs back downhill through a turbine. More recent technological advances in chemical batteries have reduced costs significantly that have pushed battery technology closer to private profitability (Going with the Flow (2014)). In the spring of 2015, Tesla Motors, an electric car manufacturer, introduced the first large scale consumer electricity storage facility that used chemical battery technology. Several studies have assessed the effects of adding bulk storage to the grid. While the focus of this literature has largely been on engineering concerns, 12 several papers have noted the environmental effects of bulk storage. For example, Lueken and Apt (2014) focuses on how bulk electricity storage affects unit commitment and electricity market arbitrage and notes that bulk storage moderately increases CO 2 emissions in the PJM market. Sioshansi (2011) estimates the effects of pairing wind generation and bulk electricity storage. Most similar to our study, Hittinger and Azevedo (2015) uses an 10 Sioshansi et al. (2010) notes that storage need not be welfare enhancing. The incentives of storage operators, generators, and consumers may not be aligned. The authors demonstrate market conditions under which incentives are properly aligned. 11 We focus here on the short-to-medium run effect of installing bulk storage. Of course, part of the incentive to encourage bulk storage is to better match wind generation that peaks overnight with demand that peaks in the late afternoon hours. Introducing inexpensive bulk storage would likely be paired with significant expansion of wind generation leading to much improved environmental performance. For this analysis, we focus on the period of time over which wind generation capacity is fixed. 12 See, e.g., Xi et al. (2014),Virasjoki et al. (2015), and Yu and Rajagopal (2015) among many others. 6

7 engineering approach to estimate the environmental effects of bulk electricity storage for 20 regions around the United States. Their results indicate that bulk storage increases emissions of CO 2 significantly and likely increases emissions of SO 2 and NO x by relatively small amounts. These estimates are identified in a slightly different way than the ones we use. 13 Additionally, we simulate the market returns of bulk storage to compare the environmental costs and the benefits of arbitrage in the electricity market in the same context, while Hittinger and Azevedo (2015) focuses exclusively on the emissions effect of bulk storage. The impact of recent natural gas prices on bulk storage has been largely unexplored. Fracking led to huge reductions in natural gas prices due to increase supply of natural gas. Holladay and LaRiviere (2016) show the commodities market did not anticipate the reduction in natural gas prices, and futures markets forecast that they will continue for the foreseeable future. Hausman and Kellogg (2015) attribute roughly $3.41/mmtbu of the recent decrease in natural gas prices to the development and adoption of fracking controlling for other potential causes such as the 2008 recession, changes in environmental regulation, or increasing investment in renewables. This decrease has affected wholesale electricity prices (Cullen and Mansur (2014), Knittel et al. (2015), and Fell et al. (2016)) and the marginal emissions from electricity generation (Holladay and LaRiviere (2016)). 14 The electricity sector is the largest single consumer of natural gas, and the current low price environment has led to significant changes in the relative costs of electricity from different fuels. These costs have led to changes in both the market price of electricity and the emissions intensity of electricity generation. As a result, there is reason to believe that changes in natural gas prices have influenced the economics of bulk storage. Our goal in this paper is to identify how natural gas price decreases impact the direct 13 Siler-Evans et al. (2012) estimates marginal emissions rates in first differences, while Holladay and LaRiviere (2016) uses fixed effects regressions. 14 These papers consider other possible changes in the energy market and policy environment that could have caused these changes including the 2008 Great Recession, changes in environmental regulation, and increased renewable penetration and suggest that the fracking boom played a large role in creating the new environment of low natural gas prices. To the extent that these or other factors played a role in the changing price of natural gas they would also have an impact on both the market and non-market benefits of bulk storage. 7

8 private and public economic gain, or losses, from bulk storage. We ignore indirect gains from forgone emissions from all wind generation and focus only on electricity dispatch shifted due to bulk storage. While bulk storage represents a large inducement to install additional wind generation, modeling that entry requires dynamic investment modeling beyond the scope of this paper. 3 A Theory of Input Prices and Bulk Storage This section presents a partial equilibrium model that extends Carson and Novan (2013). It highlights how changes in natural gas prices influence the market and direct non-market impacts of bulk storage. The key contribution of the model is that it allows for two different types of fossil fuel generation (e.g., coal and natural gas) and allows heterogeneity in the efficiency of each type of unit. The model does not produce sharp propositions but it does produce closed-form solutions for how the changes in an input price affects the market and non-market impacts from charging and discharging decisions. A natural starting place for our model is Carson and Novan (2013). That model includes both high and low demand periods and productivity heterogeneity. Specifically, their model assumes there is a total cost of production during high and low demand periods: T C(Q t ) = c(q t ) + τ e(q t ), where t indexes high and low demand periods (e.g., peak and off-peak periods), Q t indexes the quantity of generation, c(q t ) is an increase private cost function, e(q t ) is an emissions function, and τ is the social cost of the emissions from generation. In their model, c(q t ) is an increasing function that is given. The model focuses on the determinants of the cost function c(q t ) and the associated emissions function e(q t ) given a set of input prices for clean (natural gas) and dirty (coal) electricity production. Because Carson and Novan (2013) take their private and social cost functions as given, we need to add the natural gas and coal electricity production trade-off in the merit dispatch order as a function of input prices; in our model, c( ) and e( ) become a function of a price vector. Thus, there are three key elements needed in our model: clean and dirty inputs and their prices, productivity heterogeneity within fuel types, and high and low electricity demand periods. The first characteristic is the innovation over the 8

9 Carson and Novan (2013) model. The added complexity of the price vector motivates us to add some structure to the distribution of plant efficiencies and thus the function c( ). We assume that there are two types of price taking firms that can produce a homogeneous output Q. One type of firm produces Q by using a clean input x c, and the other type produces Q by using a dirty input x d. For simplicity, in each production period each firm produces a single unit of output or does not produce any output. 15 We have two periods in which demand is exogenous from the perspective of the firms and does not affect welfare in and of itself. 16 We also assume that each unit of the dirty input used in producing Q is associated with emissions that have a negative cost to the economy of size τ. 17 For simplicity we assume clean firms do not cause emissions. There is heterogeneity in the marginal productivity of both clean and dirty firms. Each clean firm c = 1,..., C has an input requirement function dictated by a single parameter θ c U[θ c, θ c ]. The parameter θ c indexes the total input required to produce a single unit of output. As a result, the firm indexed by θ c is the most efficient firm. Similarly, each dirty firm d = 1,..., D has an input requirement function dictated by a single parameter θ d U[θ d, θ d ]. Therefore, θ c and θ d index the least efficient coal and gas firms, respectively. We assume x c = θ c and x d = θ d so that efficiency maps to input requirements one-to-one. The cost of production for each firm is the product of the input requirement parameter and the exogenously determined price of their input. For example, the cost for the most efficient dirty firm to produce a unit of output is P d θ d. We assume that firms are dispatched by cost. This is consistent with firms engaging in an auction for the right to supply the market with output Q. Mirroring a competitive wholesale market in the electricity sector, the auction s outcome is that only the Q most efficient firms supply to the market and that each firm is paid a price equal to the cost of the least efficient (e.g., 15 This assumption is not particularly restrictive because we allow for heterogeneity in the firms efficiency. 16 Any welfare effects from changes in overall levels of electricity demand are second order in the context of our model. This is the primary partial equilibrium aspect of our model. 17 In this way, it is natural to interpret τ as the marginal excess damage of coal-fired electricity generation relative to natural gas. 9

10 marginal) firm. 18 More precisely, the marginal cost of production is dictated by: MC(Q) = max{p d θ d, P cθ c } s.t. Q d + Q c = Q where Q d is the total output from dirty firms, Q c is the total output from clean firms, and θd and θ c define the marginal dirty and clean firm as indexed by their productivity. If both firm types are producing then P d θd = P cθc. In the case that all type d firms are already producing, then the output cost is dictated by the most inefficient type c firm, and the marginal costs are not equalized across clean and dirty firms. 19 For example, if the price of the dirty input were sufficiently low, then the marginal cost of the least efficient dirty firm would be lower than the marginal cost of the most efficient clean firm. This situation would be represented in the model as P d θ d = P d θ d < P c θ c. In this model, the total demand for output Q in a period along with the input price vector jointly determine the input requirement of the marginal dirty and clean firms: θ d and θ c. Further, for a given price vector, the total negative cost to society associated with the dirty input is 0.5(θd + θ d)τq d. This is average dirty input of dirty firms times the cost of the externality times the total number of dirty firms producing. This particular functional form results from the uniform distribution assumption, but any distribution that maps into quantiles could be used to derive similar results. Now consider how a decrease in the price of the clean input affects the price of the output and the welfare cost of the dirty input s externality for a given Q. We assume that P c decreases from an initial value of P 0 c to P 1 c. If the initial equilibrium is indexed by zero and the new equilibrium is indexed by one, then at a clean input price of P 1 c it must be that P d θ d (0) > P 1 c θ c (0). This cannot be an equilibrium. For an interior solution with the 18 We avoid strategic bidding behavior in this model for simplicity. It is reasonable to assume that the market price is set by a trivial markup over the second highest bid. This equilibrium mirrors that of deregulated electricity markets without market power and well-functioning regulated markets. This model can be seen as a stylized version of the dispatch model in Borenstein et al. (2002) and Wolfram (1999). 19 For simplicity, our model is static with no entry or exit. The goal of the model is to highlight how changes in the input price and the heterogeneity in efficiency affect marginal emissions and the wholesale price of electricity. The question of how the changes in input prices affect entry and exit is important, but we leave it for future work. 10

11 new price vector the marginal clean firm must less efficient and the marginal dirty firm is more efficient: θc (0) < θc (1) and θd (0) > θ d (1). As a result, production from clean firms increases and production from dirty firms decreases until the marginal firm in the clean sector is less efficient than before the price change. Similarly, the marginal dirty firm would have to be more efficient than before the price decrease. There are two important effects of the new equilibrium for any given load level. First, there must be a decrease in the output price given by P 0 c θ c (0) P 1 c θ c (1). Second, there is a decrease in the total cost of the externality of 0.5τ(θd (1) + θ d (0))(Q d (1) Q d (0)). The decrease in external costs is the decrease in the average external cost of the dirty firms that stop producing multiplied by the decrease in the output of dirty firms. The first effect is the market effect of the price decrease and the second effect is the non-market effect of the price decrease. Now consider the implications of bulk storage for this model. Bulk storage alters the pattern of demand across the day by increasing demand in the low demand period (charging) and decreasing demand in the high demand period (discharging) as storage allows arbitrage. We model bulk storage as an exogenous increase in demand by q during the low demand period and a corresponding decrease in the high demand period. We evaluate bulk storage s impact on prices and emissions over stylized clean and dirty input requirement parameterizations and input price vectors. First is the stylized case where the cost of production for dirty goods is everywhere weakly lower than for clean goods: P d θ d = P c θ c. We assume there are equal numbers of clean and dirty firms and that output is uniformly distributed between zero and the total number of firms: Q U[0, Q = N c + N d ]. We also assume that each of the two output periods are random draws from the uniform distribution, with the low demand period corresponding to a random draw from the lower half of the distribution and the high demand period corresponding to a random draw from the upper half. 20 Given these price assumptions 20 These assumptions on the lower and upper half random draws mitigate the need to develop first and second order statistics for the uniform distribution that does not require them for any qualitative results. If the distribution of high load periods is skewed with a large right tail, then the price effects will likely be larger. The emissions effects are a function of the emissions intensity of the far right tail. 11

12 bulk storage increases demand for electricity when the cheaper dirty fuel is producing and decreases demand for electricity when the more expensive clean fuel is producing while keeping total demand the same over the day. $/Q Figure 1: Expensive clean input with bulk storage policy P 0 H P 1 H P 1 L P 0 L Q q L Output Low Demand Period q Q H High Demand Period Output Note: The red dashed line shows the distribution of the dirty units marginal costs and the green dotted line shows the distribution of the clean units marginal costs. The low demand period shows an increase in the output of q, all serviced by dirty units, and the high demand period shows a decrease in the output of q, all incurred by clean units because of bulk storage. Equilibrium prices increase in the low demand period and decrease in the high demand period. Figure 1 depicts the situation based on these assumptions. The red dashed line shows dirty units marginal costs and the green dotted line shows clean units marginal costs. The low demand period experiences an increase in the output of q, all serviced by dirty units, and the high demand period experiences a corresponding decrease in the output of q that clean units incur because of the bulk storage policy. Bulk storage increases prices from P 0 L to P 1 L in the low demand period and decreases them from P 0 H to P 1 H in this high demand period. We assume a fixed amount of bulk storage q in the electricity system. If the quantity of bulk storage was infinite, then firms would arbitrage price differences in the high and low demand periods to zero. The situation depicted in Figure 1 also leads to an increase in emissions since the low demand period consists entirely of dirty units. Assuming that q is very small relative to the distribution of demand for electricity, the expected non-market effect of bulk storage 12

13 over T total days is to increase emissions by 1 2 τt ( q(θ d θ d ) ). This expression is the average emission rate during charging hours times the amount of electricity charged, q, weighted by both damages and the number of charging periods T. We ignore the second order effect of an increase in emission rates caused by increased generation from charging since we assume q is small. The expression takes this exact form due to our distributional assumptions. If F (X) is a general cumulative density function of efficiency levels, we can define the marginal dirty plant efficiency at during low demand period expected load as θ = F 1 (E[Q L ]). With this definition, non-market costs increase by τt qθ with bulk storage in the more general case. In the empirical section below, we non-parametrically estimate the distribution of coal and natural gas fired power plants heat rates to verify there is significant heterogeneity in generation efficiencies. While there are non-market costs to bulk storage in this stylized case, there are market benefits from bulk storage. The market benefits stem from the more expensive clean good producing less often and the cheaper dirty good producing more often. The first order effect of bulk storage on market outcomes is to lead to arbitrage profits of (P 0 H P 0 L ) q while accounting for price adjustments leads to market arbitrage gains of (P 1 H P 1 L ) q. Total market welfare gains in the wholesale electricity market are (P 0 H Q H + P 0 L Q L) ( P 1 H (Q H q) + P 1 L (Q L q) ). We now evaluate bulk storage but consider the case where the price of the clean good falls so that the distribution of costs for clean generators is identical to that of the dirty generators. Specifically, we assume that the clean input price falls to a level where P d θ d = Pc 1 θ c and P d θ d = Pc 1 θ c. The key distinction between this stylized example and the previous one is that now there will always be both clean and dirty inputs producing at all points in time. Figure 2 shows the same bulk storage arbitrage behavior with the new price vector. Due to a lower clean input price, the orange dash-dotted line shows a mixed distribution of the marginal costs for both dirty and clean units. Both clean and dirty units serve increases in load. We test this straightforward prediction in the empirical section below by estimating changes in the dispatch of coal fired power plants as a function of their 13

14 efficiency levels (e.g. heat rates) before and after the drop in natural gas prices. Under the new price assumptions the expected non-market benefits of the above policy over T periods is zero. The reason is that clean and dirty inputs are now used equally in the generation of possible outputs. As a result, there are neither non-market gains nor non-market losses from the load shifting policy in this model so long as the distribution of efficiencies of the two technologies is identical in this specific example. There are still market gains to this for bulk storage, though, since higher cost generation is replaced with lower cost generation via storage. The market benefits of the policy are smaller in the second example than in the first case because the average difference in marginal costs between the low and high demand periods are smaller, which decrease the benefits of arbitrage. As a result the market benefits of bulk storage, while still positive, are smaller after a decrease in the price of the clean input. The lack of effect of bulk storage on non-market costs in Figure 2 is fully dependent on both the uniform distribution assumption and the cost equivalence of the most and least efficient clean and dirty units. In the earlier example, bulk storage led to increases in non-market costs because lower cost dirty units replaced higher cost clean units. As a result, bulks storage could lead to decreases in non-market costs if cheap, clean generation displaces expensive, dirty generation. 21 Similarly, a decrease in the clean input price could feasibly lead to ambiguous market benefits for the policy: if there are a large number of highly inefficient clean firms, then the marginal (dirty) firm in the high demand period has a high cost. In that case, the gains from storage will still be high after the decrease in the clean input price. Our stylized case shows the opposite situation. In general, though, if the marginal cost of clean generation 21 While determining the impacts of bulk storage on non-market and market outcomes conditional on a set of prices and production technologies is straightforward, determining the change in market and nonmarket impacts from bulk storage attributable to different aspects of the decrease in the input price of the clean good requires some attention to the proper counterfactual. With respect to market benefits, for example, when the cost of the clean input decreases there is a direct benefit of a lower priced output from the clean source through a lower marginal cost for the marginal unit even before the policy is implemented in the high demand period. Second, there is an increase in the average efficiency of dirty units due to the lower priced clean output displacing the dirty output, thereby lowering their dirty input requirement. Parsing how these two direct effects interact with the bulk storage is interesting but in this paper we are primarily concerned with the net effects we observe in the data over our sample period. 14

15 is higher than dirty generation and the clean input price decreases, the market benefits of storage, while still positive, decrease. $/Q Figure 2: Inexpensive clean input with bulk storage policy P 1 L P 0 L Q q L Output Low Demand Period P 0 H P 1 H q Q H High Demand Period Output Note: The orange dash-dotted line shows the mixed distribution of dirty units and clean units marginal costs due to the price decrease in the clean input. The low demand period shows an increase in the output of q, all serviced by a mix of dirty and clean units; and the high demand period shows a decrease in the output of q, also serviced by a mix of clean and dirty units because of bulk storage. Equilibrium prices increase in the low demand period and decrease in the high demand period. The price decreases in the high demand period are lower than before because of the decrease in the clean input price. This section characterizes change in private and direct public impacts attributable to bulk storage because of the changes in input prices. We show that the changes depend jointly on both the market and non-market costs of the generating technology is on the margin when charging and discharging decisions occur. Our stylized example shows that a decrease in the clean input price leads to smaller market benefits from bulk storage. The change in non-market impacts are ambiguous and depend on both relative prices and the distribution of the efficiency levels of clean and dirty firms. 4 Empirical analysis This section tests predictions of the theoretical model to verify its mechanisms are present in the data and any results we find in the subsequent simulations are not likely to be driven by potential confounds. To do so we use non-parametric density estimation, reduced form 15

16 econometrics, and quantile regression. We use publicly available efficiency data for large fossil fuel power plant heat rates across the U.S. We document how changes in the relative price of fuels affected power plant dispatched frequency as a function of each fuel type s productivity distribution as described by the model. In conjunction with the theoretical model, this exercise verifies the results from the simulation which follows are consistent with natural gas price changes. These results provide evidence that natural gas price decreases have changed market and non-market impacts of bulk storage rather than other change in market conditions. It also motivates our use of the closed form solutions in the simulation below which evaluate how natural gas price decreases have impacted market and non-market effects of expected bulk storage charging and discharging decisions. 4.1 Data We collect data on electricity generation, fuel inputs and pollution emissions from the EPA s Clean Air Market database. All U.S. powerplants with a nameplate capacity above 25 megawatts (MW) are required to install Continuous Emissions Monitoring Systems (CEMS) in each smokestack. These systems frequently sample the air to measure the level of pollution emissions. Electricity generating units report their hourly pollution emissions along with the quantity of electricity generated and the fuel input. 22 The results are reported to the EPA, which uses the data to ensure that power plant owners are holding pollution permits commensurate with their emissions. The data is made publicly available through EPA s Clean Air Markets Database. The data also includes time-invariant unit characteristics including capacity, fuel type and geographic location. We collect data on hourly generation, fuel input and pollution emissions as well as unit characteristics for each coal or natural gas fueled electricity generating unit that reports CEMS data to EPA for the years inclusive. 23 This time period provides a 22 An electricity generating unit is the full set of equipment required to generate electricity. It would typically consist of one or more boilers connected to one or more generators. Most power plants have more than one generating unit and these units can be brought on and off line independently. 23 We select this sample period to restrict our focus to a window of time around the natural gas price fall before generating capacity could change. After 2010 a large amount of coal capacity begins to retire from the grid. By 2012 new natural gas capacity begins to enter the market. Changes in fuel composition 16

17 sample evenly balanced across natural gas price levels with relatively constant electricity demand and a stable fuel composition of the electricity generating fleet. This creates a panel of up to 52,584 observations for each of the 3,558 electricity generating units that report CEMS data during our sample period. Units that are not generating in a particular hour do not appear in the CEMS, creating an unbalanced panel. The data set is much too large to work with directly so we aggregate the data in two ways to facilitate analysis. First, we sum across hours of the year to create a year-by-electricity generating unit sample. This creates an unbalanced panel of electricity generating units. A small number of units enter or exit during the sample and few units that operate only a few hours a year do not generate in particular sample year. 24 The final sample contains 19,958 electricity generating-unit observations. Each observation includes the electricity generating unit s total annual generation and fuel inputs as well as its reported capacity and fuel type. Using the summed generation and fuel inputs data we calculate the heat-rate, which is the standard measure of electricity generating unit efficiency, by dividing fuel inputs (measured in BTU) by electricity generation measured in kilowatt hours (kwh). This statistic maps directly to θ in the model described above. We also compared unit capacity to actual generation to calculate capacity factor, the fraction of the unit s potential capacity that utilized that year. Several model predictions relate to changes in capacity factor across fuel price levels and unit efficiency. We term this aggregation of the data the electricity generator sample and use it to test the theoretical model s predictions. We separate the sample into a high natural gas price period from and a low natural gas price period from inclusive. This approach takes advantage of the decrease in natural gas prices driven by the advent of fracking to identify the impact of input price changes on the market and non-market impacts of bulk storage. The division of the sample is consistent with Holladay and LaRiviere (2016), which finds a structural of the fleet will affect the results of the analysis. In this paper we focus on the medium term implications of the fuel price change with a relatively constant electricity generating fleet. 24 We exclude oil-fired and other fuel types (burning wood or tires for example) from this portion of the analysis. We also removed a small number of units with very high or low generation or fuel inputs levels across a year. These units typically generate a very small number of hours a year or entered incorrect data. 17

18 break in natural gas prices in January of Consistent with the theoretical model, we want to focus on the effects of input prices on bulk storage holding the existing mix of generating capacity by fuel constant to highlight short and medium run intensive margin impacts. 26 For that reason we focus on a relatively narrow window around the fall in natural gas prices, before new natural gas generation capacity comes online or existing coal capacity retires. We use the exogenous variation in natural gas prices across these two sub-samples to identify the impact of price changes in the clean input on the market and environmental impact of bulk electricity storage. The difficulty in using this approach is that our results could be attributed to any change in the electricity market, energy policy, environmental regulation or the macroeconomy that began around in January 2008 and persisted through the end of our sample. The sample period was chosen to minimize possible confounds. In related work we have explored a number of possibilities including changes in demand (up less than 0.7% across the sample), fossil fuel generation capacity, renewable generation capacity, hydroelectric dispatch and environmental regulation (Holladay and LaRiviere (2016)). In that paper we present evidence that while changes in these possible confounds occur in some regions of the country, none of these channels account for changes in marginal emission rates presented here. We believe that the change in natural gas prices is the most likely driver for our results. We present this evidence in an appendix. That said, with this empirical design we cannot conclusively demonstrate that the fracking induced change in natural gas prices alone drove these results. Readers should be aware that possible confounds may exist. In the analysis, we aggregate hourly the generation data summing within each of the three National Electric Reliability Council (NERC) region. 27 The NERC breaks the country up into three interconnections, the Eastern Interconnection, Texas (ERCOT) 25 This allows us to evaluate the environmental impact of bulk storage using marginal emissions estimates from the existing literature. 26 We view the question of how the generation stock responds to low natural gas prices and the impact on bulk storage to be an interesting, but separate question. 27 In this portion of the analysis we do not exclude any fuel types, but continue to exclude the small number of units with implausibly high or low fuel input levels across a full year. 18

19 and the Western Interconnection (WECC), which are essentially isolated from each other electrically. The Eastern Interconnection is further divided into the six regions, from north to south NPCC, MRO, RFC, SPP, SERC and FRCC. Electricity flows across NERC regions within the Eastern Interconnection are small, but non-zero. 28 Figure 3 maps the NERC regions. NERC regions represent the smallest unit of geography across which we can be reasonably certain that bulk storage capacity will be charged by electricity generating units in the same region. 29 The aggregation creates a balanced NERC region-by-hour of sample panel with 420,672 observations. 30 We term this aggregation the NERC region sample and use it in our bulk storage simulations. Figure 3: NERC regions Note: The U.S. is divided into three electrical interconnections across which very little energy flows. The Eastern Interconnection is divided into six smaller regions as well. 28 EIA data suggests that net flows of electric power across regions were on the order of 2-12 million MWhr s in 2010, with the exception of flows from the Midwest to Mid-Atlantic regions where the net flow was just over 100 million MWhrs. For perspective flows into the TRE and WECC Interconnections from the east were around 1-2 million MWhrs. 29 Graff-Zivin et al. (2014) adopts a similar approach to evaluating the environmental impacts of electrical vehicles. 30 That is 24 hours a day, 365 days a year for six years, plus a leap year day for observations for each of the eight NERC regions. 19

20 4.2 Empirical analysis We begin our empirical analysis by estimating how the relationship between electricity generating unit dispatch and efficiency were affected by the change in natural gas prices. In line with the theoretical model, we pay careful attention to describing heterogeneity in power plant heat rates and how a particular plant s dispatch decision relates to their heat rate before and after natural gas price changes. We employ the electricity generating unit sample and not the hourly aggregated data for this portion of the analysis. Table 1 presents summary statistics from this sample by fuel type and natural gas price regime. Coal fueled electricity generating units generate more electricity, use more fuel input, have a lower heat rate (meaning they are more efficient), a larger capacity and higher nameplate capacity factor compared to natural gas fueled electricity generating units. Across the two fuel price level sub-samples generation falls at coal units and increases at natural gas units when natural gas prices fall. Average heat rate improves significantly at coal plants as less productive units generate less, raising the average productivity. Average heat rate decreases slightly at natural gas units and capacity factor increases slightly. The decrease in heat rate (increased efficiency) is somewhat surprising as the fall in natural gas prices induced more generation from less efficient combustion turbines, but that increase is offset by an even larger increase in generation at the most efficient units, which had significant unused capacity during the high natural gas price portion of the sample. Figure 4 displays the distribution of electricity generating units by fuel type for coal and natural gas calculated from the CEMS data s hourly fuel input and electricity output (e.g., (θ d, θ d ) and (θ c, θ c ) from the theoretical model). There is significant variation in the productivity within and across fuels. On average coal plants are more productive than natural gas plants and there is less variation in productivity among coal plants. Natural gas plants are, on average, less productive than coal plants, but the distribution is doublepeaked. The higher productivity, smaller, peak is a set of plants that employ combined cycle generation technology which uses waste heat to generate additional power. The other peak represents natural gas fueled units that employ older technologies, typically combustion turbines that do not recycle waste heat. 20

21 Table 1: Electricity generating unit summary statistics Coal Fueled Units Nameplate Cap. (MW) (268.6) (269.8) Generation (GWh) (1912.2) (1904.1) Heat Input (billions of BTU) ( ) ( ) Heat Rate (932.5) (858.6) Capacity Factor (0.152) (0.191) Observations 5498 Natural Gas Fueled Units Nameplate Cap. (MW) (177.8) (175.0) Generation (GWh) (440.1) (462.8) Heat Input (billions of BTU) (3824.6) (3815.6) Heat Rate (2285.6) (2429.5) Capacity Factor (0.159) (0.166) Observations Note: Sample statistics for coal fired electricity generating units (top panel) and natural gas fired electricity generating units (bottom panel). Table reports sample means with standard deviations in parentheses. Left column reports averages for the high natural gas price portion of the sample and the right column for the low gas price portion. Heat input is the energy content of fuel consumed by the unit measured in billions of BTU s. Heat rate reports the nameplate capacity weighted average of each unit s heat rate, measured as fuel input (in BTU) over electricity generation (in KWhr). Capacity factor is nameplate capacity weighted average of the fraction of the unit s maximum generation that is actually dispatched. 21

22 Figure 4: Distribution of power plant productivity by fuel type kdensity Heat Rate Gas Coal Note: Nameplate capacity weighted density plots of the relative productivity of power plants using the dirty (coal) and clean (natural gas) input aggregated from Productivity is measured here as heat rate, which is defined as fuel input/electricity output and is the typical measure of a powerplants efficiency. Coal plants are relatively tightly clustered, but natural gas plants rely on two different technologies. High productivity combined-cycle plants use waste heat to generate additional electricity. Less efficient turbine plants use only natural gas fuel and typically do not recycle their waste heat for electricity generation. 22

23 The theory model predicts that production from clean (natural gas fired) units will increase and production from dirty (coal fired) units will decrease until the marginal firm in the clean sector is less efficient than before the price change. At the same time, the marginal dirty firm would have to be more efficient than before the price decrease. We now estimate a series of regressions that take advantage of the natural gas price decrease to test for changes in capacity factor. The regressions take the form: CapF actor jt = α + βlog(heatrate jt ) + ηgas jt + γ1{t >= 2008} t + φx jt + ɛ jt, (1) where j indexes electricity generating unit and t denotes year. CapF actor is the portion of a unit s capacity employed over the course of the year ranging from 0 to 1. HeatRate is the unit s efficiency over the course of the year measured as heat input over electricity generation output. Gas is an indicator for fuel type, equal to 0 coal and 1 for natural gas. 1{t >= 2008} is an indicator for the low natural gas price portion of the sample that equals 0 for and 1 for X is a matrix of unit characteristics including capacity, an indicator for whether the plant can employ multiple fuels and the age of the unit in decades. We also interact the heat rate, gas and low gas price portion of the sample indicators to evaluate the changing impact of unit efficiency across fuel types and gas price levels, which the model suggests is an important determinant of the environmental impact of bulk storage. Table 2 summarizes the results of these regressions. Column 1 reports the elasticity of capacity factor with respect to heat rate. A one percent increase in heat rate is associated with a 0.55% decrease in capacity factor. Column 2 adds a set of electricity generating unit controls. The controls are have the expected signs, older plants have a lower capacity factor and larger plants have a higher capacity factor. Natural gas units have significantly lower capacity factors. Column 3 interacts heat rate with the low natural gas price indicator. There is no significant difference in the impact of unit productivity on capacity factor across the high and low natural gas price portions of the sample. Column 4 interacts fuel type and heat rate. Natural gas units are around four percent more sensitive to heat rate than coal units. 23

24 Table 2: : Capacity factor and heat rate across fuel prices (1) (2) (3) (4) (5) Cap. Factor Cap. Factor Cap. Factor Cap. Factor Cap. Factor Log(Heat Rate) *** *** (0.012) (0.009) Indicator Log(Heat Rate) *** (0.012) Indicator Log(Heat Rate) *** (0.012) Coal Indicator Log(Heat Rate) *** (0.009) Natural Gas Indicator Log(Heat Rate) *** (0.009) Ind. Coal Ind. Log(Heat Rate) *** (0.009) Ind. Coal Ind. Log(Heat Rate) *** (0.009) Ind. Natural Gas Ind. Log(Heat Rate) *** (0.009) Ind. Natural Gas Ind. Log(Heat Rate) *** (0.009) Natural Gas Indicator *** (0.005) Age (decades) ** *** 0.002** ** (0.001) (0.001) (0.001) (0.001) Multiple Fuels *** (0.003) (0.004) (0.003) (0.003) Nameplate Capacity (GW) 0.169*** 0.389*** 0.172*** 0.172*** (0.007) (0.013) (0.007) (0.007) Constant 5.435*** 2.626*** *** 2.211*** 2.173*** (0.109) (0.235) (0.280) (0.237) (0.234) R 2 N Note: Results of regressions exploring the relationship between electric generating unit efficiency (heat rate) and dispatch intensity (capacity factor). Each column reports the results of a regression with 17,272 observations. Heat rate is calculated as fuel inputs over electricity generation so lower heat rates indicate more efficiency units. Age is measured in decades to enhance readability. Natural gas and coal indicators are binary variables that equal 1 for a particular fuel type and indicators represent the high and low gas price portions of the sample respectively. Newey West standard errors reported in parenthesis to correct for potential serial correlation. *** significant at the 1% level, ** significant at the 5% level, * significant at the 1% level. 24

25 Finally, column 5 reports a triple interaction of heat rate, fuel type and an indicator for the low natural gas price portion of the sample. The coefficients allow us to evaluate the sensitivity of coal and gas plants capacity factor to heat rate across fuel price levels. This specification highlights the importance of considering changing fuel prices and the distribution of unit efficiency jointly when assessing changes in the way electric generators are dispatched. Coal fueled units capacity factor become almost one percent more sensitive to heat rate after the fall in natural gas prices. The difference between those coefficients is statistically significant at the one percent level. This suggests that all else equal a plant with a one standard deviation higher heat rate would expect a capacity factor just over ten percent above a less efficient plant during the low natural gas price portion of the sample. 31 Natural gas units capacity factors sensitivity to natural gas prices is essentially unchanged. 32 This is consistent with the prediction that inefficient coal units become more exposed to natural gas competition in the low natural gas price portion of the sample. The model predicts that the impact of cheap natural will not be uniform, but will primarily affect units near the margin. Reduced natural gas prices will shift formerly marginal electricity generating units into higher levels of utilization. Coal units that were near the margin are likely to be displaced by the higher level of natural gas generation. For that reason estimates across the sample may hide the variation predicted by the simple model. To identify the impact of the reduction in natural gas prices across the electricity generating unit distribution we estimate a series of quantile regressions to analyze the impact of changing natural gas prices across the capacity factor distribution. The results are presented in table 3. Quantile regressions use the full sample to estimate marginal effects at different points on the distribution of the dependent variable. They are less sensitive to outliers, but more importantly for this application, they allow us to assess 31 Based on the summary statistics from table 1 one standard deviation is around 9% of the mean. The average coal plant capacity factor during the low natural gas price portion of the sample is When separating the effects on combustion turbines and combined cycle plants we find small, but statistically significant increases in sensitivity to heat rate at combustion turbines and decreases at combined cycle plants. 25

26 whether the marginal effect of low natural gas price varies across levels of capacity factor. The quantile estimates are simple regressions of capacity factor on heat rate and fuel price level indicators at different points of the capacity factor distribution, estimated jointly. We restrict the sample to a single fuel type and estimate CapF actor jt = α(τ) + β(τ)heatrate jt + γ(τ)1{t >= 2008} t + ɛ jt, for τ=0.1, 0.25, 0.5, 0.75 and 0.9 quantiles. The γ coefficients are the change in capacity factor during the low natural gas price portion of the sample at that point in the capacity factor distribution. Table 3: Quantile regressions for unit capacity factor and heat rate across gas price levels Coal Fired Units τ=0.1 τ=0.25 τ=0.5 τ=0.75 τ=0.9 Log(Heat Rate) ** ** ** ** *** (0.067) (0.053) (0.033) (0.049) (0.059) Indicator ** ** ** ** *** (0.011) (0.007) (0.006) (0.005) (0.007) Constant 8.732** 6.919** 4.470** 3.192** 3.032*** (0.609) (0.486) (0.300) (0.450) (0.545) Pseudo-R Natural Gas Fired Units τ=0.1 τ=0.25 τ=0.5 τ=0.75 τ=0.9 Log(Heat Rate) ** ** ** ** *** (0.001) (0.002) (0.007) (0.008) (0.016) Indicator ** ** ** ** (0.000) (0.000) (0.001) (0.003) (0.009) Constant 0.133** 0.433** 1.306** 2.781** 2.910*** (0.009) (0.022) (0.065) (0.078) (0.154) Pseudo-R Note: Each panel reports a series of quantile regressions with electricity generating unit capacity factor as the dependent variable and log heat rate and an indicator for years as the dependent variables. The top panel reports results for a sample restricted to coal fired units with X observations. The bottom panel reports results for a sample restricted to natural gas fired generators with Y observations. The indicator equals 1 for years and is a proxy for the low natural gas price era. Heat rate is measured as fuel input/electricity output so high heat units are less efficient. All standard errors clustered at the unit level. *** significant at the 1% level, ** significant at the 5% level, * significant at the 1% level. The results suggest that high capacity factor natural gas units are used relatively more during the low natural gas price era. Each column of table 3 reports results for a different quantile of the capacity factor distribution. The top panel reports results for coal fueled units and the bottom panel reports results for natural gas fired units. Coal fueled electricity generating units see their 26

27 Figure 5: Productivity and output by input type and price Coal Gas Capacity Factor Heat Rate Capacity Factor Heat Rate Note: Productivity, measured by heat rate, and capacity factor for coal (left panel) and natural gas (right panel). Solid line represents the sample period with relatively high natural gas prices and dotted line represents the low gas price portion of the sample. The fall in natural gas price led to increases in output in relatively productive gas plants and decreases statistically significant decreases in output at moderate and low productivity coal plants. Note the vertical axis is different across the two graphs to highlight the relative difference across gas price regimes, rather than the difference in capacity factors across fuels. capacity factor fall in each of the quantiles. The decrease in capacity factor is largest for the lowest capacity factor units. This implies that the least utilized coal units see the largest reductions in generation after the fall in natural gas prices. For natural gas units there is also a small reduction in capacity factor, but reductions are extremely small, do not present a clear pattern across capacity factors and are less precisely estimated. Figure 5 presents univariate kernel regressions of capacity factor on heat rate. 33 We run separate smoothing regressions for both coal and natural gas in each natural gas price portion of the sample. The left panel presents the results for coal and the right for natural gas. In each panel the solid line represents (the expensive natural gas portion of the sample) and the dotted line represents (the cheap natural gas portion). Coal capacity factor is significantly lower for all but the most productive (lowest heat rate) plants. 34 While the most productive coal plants are at or near the maximum achievable 33 The regressions equations are: CapF actor jt = α jt + βheatrate jt on a sample restricted to a single fuel type and price level time span. They are estimated with an Epanechnikov kernel smoothing function. 34 Beyond heat rates of around 14,000, which is the 99th percentile, estimates become extremely imprecise, but the point estimates from the low natural gas price portion of the sample remain lower. 27

28 capacity factor in both gas price regimes, less efficient coal units have untapped capacity at all natural gas prices. Natural gas capacity factor rises for the most productive (lowest heat rate) plants. 35 The reduction in output at less productive coal plants and the increase in output at less productive gas plants are both consistent with the model: when gas prices decrease, average generation by coal decreases (e.g., coal has a lower capacity factor). 36 The kernel regressions reported here cannot control for any covariates. To explore how adding additional controls affected the nonlinear relationship identified in Figure 5, we estimated a regression that bucketed generators into deciles of heat rate by fuel type and interacted those deciles with heat rate, the indicator and fuel type. 37 We also included the controls reported in quantile regressions. The results are available in the appendix and broadly consistent with those reported in Figure 5. The empirical results highlight the impact of the fall in natural gas prices on electricity generating unit capacity factor by fuel type. These results suggest that changes in fuel prices will have an impact on both the market and non-market returns of bulk electricity storage through the change in the types of unit on the margin. While we believe that changes in natural gas prices are the most likely drivers of the results presented here, we cannot rule out all alternative channels. Other factors that affect the grid and differ systematically across the two segments of our sample could also be affecting the results Bulk Electricity Storage Simulations In this section we simulate the change in market and non-market impacts of installing a small amount of bulk electricity storage before and after the natural gas price decrease. 35 There are a small number of observations of very low heat rate units with very low capacity factors. Fitting these observations generates the positive slope portion of the natural gas capacity factor curves. Restricting the sample to heat rates above the first percentile eliminates that region of positive slope. Most of these high productivity, low capacity factor units are relatively small gas units at larger coal plants. 36 Running the same kernel regressions weighting observations by nameplate capacity does not materially affect the results. For coal the confidence intervals overlap for the highest heat rate (least efficient) units. For natural gas the gap between fuel price levels is slightly larger for the lowest heat rate units. 37 We thank an anonymous referee for suggesting this approach. 38 We note that that demand, the fuel mix of the generating fleet and renewable generating capacity are stable during our sample period. We are also note that changes in pollution permit prices, driven in part by the fall in natural gas prices, make coal more competitive and mute the effects we find here. 28

29 This is the main contribution of the paper as we operationalize the insights of the theoretical model and empirical exercise consistent with it above. We proceed in two ways. First, we simulate the market and non-market impacts charging natural gas prices for each of the eight NERC regions in the U.S. For this exercise we use averaged on-peak and off-peak aggregate prices. This provides insight into heterogeneity in the charging and discharging decisions over space. Second, we investigate the direct market impacts more carefully using hourly price data from PJM. This sacrifices external validity for better internal validity due to the higher frequency of the electricity price data. The results are qualitatively consistent across the two approaches. 5.1 Nationwide simulation For our NERC level analysis we use an aggregation of the raw generation data which we term the hourly data set. We aggregate all generation in a NERC region for each hour of our sample from This generates a balanced panel of 420,672 hourly generation and fuel input observations. We use this data set to implement our nationwide bulk storage simulation. 39 We simulate installing a small amount of high capacity electricity storage in each NERC region at the start of the study period. Recall that bulk storage technologies differ in their capacity and charging rates, we abstract from this issue and follow Carson and Novan (2013) in assuming that storage can charge and discharge completely in one hour. We further assume that there are no transmission constraints on the system so that storage anywhere in a NERC region can serve demand anywhere else in the region. 40 In the absence of detailed price data for NERC regions across the country we also assume that the storage capacity is charged during the lowest demand hour of the day and discharged during the highest demand hour. This strategy would maximize the market returns to 39 We also use this data to demonstrate that the observed changes in prices and emissions rates were not driven by changes in demand. Hourly demand across our two time periods is statistically unchanged. The results of this analysis are available upon request. 40 Several papers employing an engineering approach highlight the importance of storage location and transmission constraints in evaluating the market returns to bulk storage. See Walawalkar et al. (2007) and Sioshansi et al. (2009) for examples. 29

30 the storage operator. Note that this approach does not allow for trading across NERC regions. While flows of electricity across regions are small, Graff-Zivin et al. (2014) notes that they are non-zero. Unfortunately, in the absence of price data we are unable to accurately model which region would an import and which region would be an exporter, so we implement the simulation restricting storage to be charged and discharged in the same region. Holladay and LaRiviere (2016) report marginal emissions by hour of the day, month of year and fuel price regime. Using these estimated marginal emissions we calculate the additional emissions that occur during the charging process and the emissions avoided during the discharge process. We then aggregate the emissions impacts for the high and low natural gas price regimes across each NERC region. We use the aggregate generation data to identify the lowest and highest generation hours of the day, for each day of the sample. We designate those are the charging and discharging hours, respectively. We use estimated CO 2 marginal emissions rates from Holladay and LaRiviere (2016) to determine the emissions generated during charging hours and averted during discharging hours. 41 We calculate: Storage h,m,r,n = ME h,m,r,n if Gen h,m,r,n = Min d (Gen h,m,r,n ) ME h,m,r,n if Gen h,m,r,n = Max d (Gen h,m,r,n ) 0 otherwise, (2) where Storage h,m,r,n is the CO 2 emissions impact of adding a single MW of bulk storage to the grid in hour h, of month of year m, during gas price regime r in NERC region n. ME h,m,r,n is the estimated marginal emissions. Gen h,m,r,n represents the hourly generation measured in MW s and Min d and Max d represent the day of sample maximum and minimum load in an hour, gas price regime and NERC region. 41 We employ estimated marginal emissions rates from the literature rather than the actual emissions rates in those particular hours to capture the average effect of storage. This means that individual hours with extreme levels of emissions may be poorly modeled by that average. Since our aim is to predict the general effect of bulk storage across price regimes we believe this is a reasonable tradeoff. 30

31 In the simulation, storage is typically charged overnight, most commonly hours 2-4 of the day (e.g., 1am to 3am), and discharged over late afternoon hours which vary between depending on the season. Table 4 describes the non-market (CO 2 emissions) and market (electricity price arbitrage) outcomes of installing bulk storage in each NERC region. Columns 1 and 2 report the emissions associated with charging (ME hmr Charge h ) and discharging (ME hmr Discharge h ) a MW of bulk storage capacity by natural gas price regime. Marginal emissions rates are higher overnight when coal is on the margin and lower during the afternoon when natural gas tends to be on the margin, meaning bulk storage leads to net increases in emissions. Charging becomes slightly less dirty and discharging prevents slightly more emissions in the low natural gas price portion of the sample. We monetize changes in CO 2 emissions at $36 per ton, the value that U.S. Environmental Protection Agency uses when conducting cost benefit analysis of new regulations. 42 We also approximate the electricity market returns associated with charging and discharging per MW of electricity stored. We collect annual average prices for offpeak and onpeak electricity at hubs around the country from the Federal Energy Regulatory Commissions (FERC) Electric Market Overview for each year in our simulation. 43 The overview reports average prices by peak type for twenty-six trading hubs around the country. We assign each hub to a NERC region and average across all hubs in a region. These prices represent weighted averages across hours within a peak type. It is important to note that storage operators will pay less to charge and get paid more to discharge than the average cost across a peak type. That implies that these prices represent a lower bound on the market returns to operating bulk storage. 44 The external (environmental) costs associated with charging are reported in columns 1 42 This figure represents the costs of CO 2 emissions in 2015 estimated with 2.5% discount rate. See Technical Update of the Social Cost of Carbon for Regulatory Impact Analysis, as Revised in July 2015 for more details. 43 The FERC Overviews are available here: We collected price data from the Jan 2009 and Jan 2011 reports, but the same pricing data appears in various months. 44 In section 5.2 we take advantage of hourly pricing data for different parts of a single electricity market to address this issue. 31

32 (high gas prices) and 2 (low gas prices). 45 Charging generates additional emissions in each NERC region because it substitutes pollution intensive coal serving baseload overnight for relatively clean natural gas serving peak load during high demand hours. The external costs are highest in MRO and SPP which have extensive coal generating capacity. External costs are lowest in NPCC and WECC, regions with relatively little coal capacity. More importantly for our application, the level of cross-region variation in external costs is essentially unchanged by the shift in relative fuel prices, but that hides significant churn. FRCC sees a significant increase in emissions associated with bulk storage during the high natural gas price portion of the sample. TRE and WECC experience smaller increases in emissions from bulk storage. SERC and RFC, on the other hand, see big decreases in emissions. The changes in emissions impacts across natural gas price levels are associated with the types of generation capacity installed in each region. FRCC, TRE and WECC each have relatively little coal and significant combined cycle gas resources. SERC and RFC have very high coal capacity and much lower levels of natural gas and therefore are less exposed to the relative price change. In no region does the social costs of storage become negative. A negative social cost would imply there is a social benefit, in terms of carbon reduction, to bulk storage. Put another way, even in the low natural gas price regime, there is never a social motive to subsidize bulk storage. This is consistent with the findings in Carson and Novan (2013) despite observed differences in the dispatch order both on and off peak. There are two caveats. First, we only examine data through the end of While input prices have not changed a great deal since then, it is possible the installed capacity, and therefore the efficiency distribution has. Second, our estimates reflect only the CO 2 emissions impact of bulk storage. We do not account for other co-benefits. Importantly, though, this methodology can be used at any point in time and for any region to calculate 45 The environmental costs of bulk storage are calculated using only CO 2 emissions. Bulk storage would be associated with changes in SO 2 and NO x emissions as well, but those emissions are regulated in capand-trade systems over a large portion of the country. Bulk storage in regulated jurisdictions would change the price of permits, but not the total level of emissions if the cap continues to bind. For that reason we choose to focus on CO 2 emissions, which are uncapped nationwide. 32

33 the non-market impacts of storage. We implement the same procedure in the simulation for PJM using higher frequency data be]low as an example. The market returns to arbitrage are reported in columns 3 (high gas prices) and 4 (low gas prices). The private returns to charging are the revenue from dispatching storage during the onpeak portion of the day minus the costs of charging offpeak. to inexpensive natural gas greatly reduces the private returns to storage. The move Natural gas generators were on the margin during the high natural gas price portion of the sampel leading to high electricity prices and returns to discharging storage. After the fall in natural gas prices peak electricity prices fall considerably reducing the returns to discharging. Inexpensive natural gas also greatly reduces the cross-region variation in private returns to storage. Columns 5 and 6 report the net social returns to bulk storage including both external costs and private returns. The private returns to bulk storage are an order of magnitude larger than the external costs suggesting that bulk storage is welfare enhancing despite the associated increase in emissions. The reduction in natural gas prices is associated with big reductions in private benefits and relatively small changes in external costs reducing the welfare benefits of bulk storage. The results show significant variation in the private benefits and external costs across both NERC regions and natural gas price regimes. Table 5 below describes the generating capacity in each NERC region by fuel type. Regions with relatively large fractions of coal and nuclear generation, traditional base load fuels, generally enjoy the largest private benefits from bulk storage. After the fall in natural gas prices these same regions experience the biggest reduction in benefits. Regions with significant natural gas capacity, particularly efficient combined cycle units, get the smallest benefits from bulk storage. Regions with large natural gas capacity during the high natural gas price regime would have benefited greatly from bulk storage. Storage allows intertemporal substitution of inexpensive coal and nuclear generation for high price natural gas generation. Regions with large installed coal and nuclear bases relied less heavily on expensive natural gas to serve peak load and enjoyed smaller benefits. Because coal generation is significantly more pol- 33

34 Table 4: Private benefits and external costs of bulk electricity storage (Monthly Marginal Emissions) Non-market Costs Market Benefits Net Benefits High Low High Low High Low (1) (2) (3) (4) (5) (6) FRCC MRO NPCC RFC SERC SPP TRE WECC Note: All numbers represent average daily $/MWh. The left two columns report the per megawatt external costs of additional CO 2 emissions associated with bulk electricity storage estimated monthly. The middle two columns describe the net private benefits of bulk storage per megawatt in the high and low natural gas price regimes respectively. The right two columns report the net benefits of bulk electricity storage per megawatt. The private benefits are calculated from the average spot price of electricity during off- and on-peak hours by NERC region calculated from data reported in FERC Electric Market Reports: National Overview and represent a lower bound on the actual private benefits. lution intensive the impact of storage on emissions is reversed. Facilitating the substitution of relatively dirty coal generation for clean natural gas increases total emissions. These results are consistent with the predictions of the model. Production with the clean input (natural gas) is relatively less efficient than production from the dirty input (coal) meaning that production levels are lower for plants using the clean input. A decrease in the price of the clean input shifts production from dirty generators and lowers output prices disproportionately in high demand periods. This in turn reduces the market benefits of the bulk storage technology that shifts production from high to low demand hours. At the same time the non-market costs of shifting production are reduced to the extent the clean input has displaced the dirty input in low demand periods. 5.2 Bulk Storage in PJM Using on- and off-peak average prices to estimate the private returns of bulk storage produces estimates that are a lower bound on the benefit of bulk storage. To address that concern, we now employ the same simulation approach focusing on the PJM wholesale 34

35 Table 5: Generating Capacity by Fuel Type and NERC Region NERC Coal Oil CC Gas Other Gas Nuclear Other FRCC MRO NPCC RFC SERC SPP TRE WECC Note: Data on generating capacity by fuel type as a percent of total capacity in NERC region. CC gas are efficient combined cycle natural gas fired units. Other gas includes various combustion turbines with lower levels of efficiency. Source: EPA Egrid for 2005 plant and unit reports. electricity market. PJM provides hourly price data that does not exist for regions outside of wholesale markets. Using hourly data we can directly identify which hours would be best for charging and discharging for each day in our sample rather than relying on onpeak and off peak definitions as in the previous section. To implement this simulation we collected a separate dataset consisting of hourly wholesale prices from July of 2005 through the end of The data includes demand and prices for PJM, whose three regions are mapped in Figure 6. The South region consists primarily of Virginia, the Mid-Atlantic region stretches includes Maryland, Delaware, New Jersey and most of Pennsylvania, and the Western region includes West Virginia, Ohio and parts of Illinois. Figure 6 describes the area served by PJM. The wholesale price for PJM is a load weighted average of prices at approximately 16,000 price nodes inside the PJM footprint. Figure 7 reports the average hourly real time price for PJM across fuel price regimes. 47 The changes in real time price across the high and low gas price regimes are significant. The average (unweighted) real time price 46 Historical data on wholesale prices and load are only available beginning in July of We select this sample period to restrict our focus to a relatively constant PJM service territory. In 2011 PJM added FirstEnergy, which includes much of the rest of Ohio and portions of Michigan. During our sample period the only expansion was the addition of Dominion in Virginia in May of Restricting our analysis to focus on June 2005 on does not materially affect the results. 47 We also collect hourly load data for PJM over the sample period. Load increases by approximately 0.2%. This suggests that observed changes in emissions rates are driven by re-ordering of fuel on the supply curve rather than a change in demand. 35

36 Figure 6: PJM Region Map for wholesale electricity falls by more than 30%. The reduction is largest over the peak hours. Prices between 8-9 PM fall from to 49.50, a reduction of 54%. Over the early morning hours average price reductions range from 10% to 15%. Because the largest reductions in price happen in the hours with the highest load, the load weighted average price falls by 42%, more than the unweighted average price reduction. In the appendix we report average hourly real time load and price for each month of the year. In the appendix we present evidence that demand has not changed across the high and low natural gas portions of the sample. Using the PJM data described above, we find the lowest and highest hourly wholesale price for each hour of the day during the sample period. We denote these hours as the charging and discharging hours respectively. We then implement the same simulation procedure described above to estimate both the market and environmental impacts of bulk storage in PJM. In the simulation, storage is typically charged in the overnight hours, most commonly hours 2-4 of the day, and discharged over late afternoon hours which vary between depending on the season. Table 6 describes the impacts of 36