Marginal emissions factors and marginal impact factors: a tool for policy assessment Ines Azevedo Associate Professor Department of Engineering and Public Policy Carnegie Mellon University Co-Director Climate and Energy Decision Making Center

How do different interventions affect the emissions and damages from the U.S. electric grid? Several papers in: PNAS, ES&T, ERL, Applied Energy.

Are we helping the environment more by increasing solar in California or in Pennsylvania?

Solar PV - The locations that provide the largest electricity output are not the ones that have the largest climate, health, and environmental benefits. Energy Performance Avoided CO2 per kW (kg & $) Health and environmental benefits This tells us that, if you’re goal is mitigating climate change, the best place for a solar panel in Kansas, Nebraska, or the Dakotas, where there’s a moderate solar resources, but you’re primarily displacing carbon-intensive coal plants. In California or Arizona, gas-fired generators are predominantly on the margin and as a result, solar panels displace relatively little CO2 emissions. There’s a couple of surprises here. Despite the poor solar resource, a solar panel in South Carilna is expected to displace 20% more CO2 emissions than a panel in Arizona. Ohio and Pennsylvania are better than southern California, and in terms of CO2 emissions, Florida is about the worst place for a solar panel. References: (1) Siler-Evans, K., Azevedo, I. L., Morgan, M.G, Apt, J. (2013). Regional variations in the health, environmental, and climate benefits from wind and solar generation, Proceedings of the National Academy of Sciences, 110 (29), 11768-11773; (2) Siler-Evans. K., Azevedo, I.L., Morgan, M.G., (2012). Marginal emissions factors for the US electricity system. Environmental Science & Technology, 46 (9): 4742–4748.

What are the CO2 life-cycle emissions from light duty passenger gasoline and plug-in electric vehicles across the U.S.? (or… Are we helping de-carbonization more if we choose an electric car or an gasoline car?)

gCO2eq mi−1

eNissan Leaf > epriusHEV coal-heavy electricity grid rural counties (highway driving cycle) cold weather eVolt> epriusHEV Volt consumes + gasoline per mile in charge-sustaining mode (after the battery is depleted) than the Prius HEV, and it consumes + electricity per mile than the Leaf in charge-depleting (CD) mode (when the battery is charged) at high temperatures. Further, in cold weather the Volt consumes both gasoline and electricity in CD mode. PHEV Prius consumes less gasoline than the HEV Prius in city driving conditions and more gasoline than the HEV Prius in highway driving conditions. gCO2eq mi−1

gCO2eq mi−1

We have now many pieces using the marginal emissions and damages factors… Gingerich, D., Sun, X., Behrer, P., Azevedo, I.L., Mauter, M., (2017). Air emissions implications of expanded wastewater treatment at coal-fired generators. Proceedings of the National Academy of Sciences, published ahead of print February 6, 2017 Keen, J., Apt, L. (2016). Are high penetrations of commercial cogeneration good for society?. Environmental Research Letters 11.12. Yuksel, T., Tamayao, M., Hendrickson, C., Azevedo, I.L., Michalek, J., (2016). Effect of regional grid mix, driving patterns and climate on the comparative carbon footprint of gasoline and plug-in electric vehicles in the United States. Environmental Research Letters, 11. Tamayao, M. , Michalek, J., Hendrickson, C., Azevedo I.L., (2015). Regional variability and uncertainty of electric vehicle life cycle CO2 emissions across the United States. Environmental Science & Technology, 49 (14). Hittinger, E., Azevedo, I.L., (2015). Bulk energy storage increases US electricity system emissions. Environmental Science & Technology, 49 (5). Gilbraith, N., Azevedo, I.L., Jaramillo, P., (2014). Regional energy and GHG savings from building codes across the United States. Environmental Science & Technology, 48 (24). Siler-Evans, K., Azevedo, I.L., Morgan, M.G, Apt, J. (2013). Regional variations in the health, environmental, and climate benefits from wind and solar generation. Proceedings of the National Academy of Sciences, 110 (29), 11768-11773. Siler-Evans, K., Azevedo, I.L., Morgan, M.G., (2012). Marginal emissions factors for the US electricity system. Environmental Science & Technology, 46 (9), 4742–4748.

The time at which electric vehicles are charged will affect whether they are reducing or increasing carbon emissions. So if you invest in increasing renewables in the grid, or if you invest in an energy efficiency measure and you save some amount of electricity, that means that some power plant in the system will scale back a little bit, and we should see a reduction in emissions. In an ideal world, we would know exactly what generators would be scaled back for any location in the United States. For example, in the top figure I am showing a made up dispatch curve. The dispatch curve shows the order in which the generators are dispatch to meet demand. In the lower picture, I show what the associated emissions per unit of electricity generated are associated with each of these generators. Now, what we would like to track is how often does an energy efficiency or renewable intervention reduce each of these generators – which will depend of the level of demand at each instant. This is very difficult to do for all regions of the US using dispatch models. Marginal Emissions Factors help us understand the relationship between these two pieces. Figures from Azevedo – this is a schematic only, it does not represent a real system

∆generation ∆emissions So if you invest in increasing renewables in the grid, or if you invest in an energy efficiency measure and you save some amount of electricity, that means that some power plant in the system will scale back a little bit, and we should see a reduction in emissions. In an ideal world, we would know exactly what generators would be scaled back for any location in the United States. For example, in the top figure I am showing a made up dispatch curve. The dispatch curve shows the order in which the generators are dispatch to meet demand. In the lower picture, I show what the associated emissions per unit of electricity generated are associated with each of these generators. Now, what we would like to track is how often does an energy efficiency or renewable intervention reduce each of these generators – which will depend of the level of demand at each instant. This is very difficult to do for all regions of the US using dispatch models. Marginal Emissions Factors help us understand the relationship between these two pieces. ∆emissions Figures from Azevedo – this is a schematic only, it does not represent a real system

Estimating marginal emissions For every fossil fuel power plant we have hourly measured emissions of CO2, SO2, NOX at the stack (CEMS data from the EPA) For PM2.5 we use NEI annual data, and use the correlation with SO2 emission to estimate hourly emissions. For each region (state, balancing area, eGrid, NERC), we run separate regressions by time of day and season: ∆emissionsregion,h,season [ton CO2/h] = α∆generationregion,h,season [MWh] + ε  Results from the APEEP model give the average, dollar-per-ton damages for each pollutant (SO2, NOx, PM2.5, PM10, VOCs, and NH3) emitted by point sources. In APEEP, plants with an effective stack height higher than 500m are modeled individually for each facility, so the damages (in $/ton) from APEEP from those plants are used directly in our model. We use this assumption for 656 power plants. Plants with stack heights of less than 500m are modeled by APEEP as a county level emissions source (a point source in the middle of the county with either a low or medium effective stack height). For xyz power plants, we assumed that the plants have the average county-level point source damages (in $/ton of pollutant) estimates from APEEP for a specific county and for a specific stack height. For each source location, APEEP uses a Gaussian plume model to estimate the dispersion of emissions and the resulting concentrations in each county. Dose-response functions are used to estimate physical effects of affected populations and other receptors (crops, forests, materials, etc.). Physical effects are translated to monetary values using market prices for lost commodities, costs of illnesses, a value of a statistical life (VSL), and other non-market valuations from the literature. VOCs, NH3, and PM10 are excluded from this analysis because they result in damages that are, on average, more than two orders of magnitude lower than damages from other pollutants. At this point in the process we have hourly damages from individual power plants. The next step is to sum them up to find the houlry damages from all power plants in a region. And we do this separately for each of the 22 egrid subregions shown here. In this analysis we make the assumption that regions are completely isolated… we ignore imports and exports of electricity between regions. We then bin the data in 5% increments of gross generation. For each of these 20 gross generation bins, we regress the hourly change in damages per pollutant as function of the change in generations. We show a sample result for Texas. At high demand marginal CO2 and SO2 rates decrease as gas accounts for a larger share of marginal generation. In this case, the marginal NOx rates increase with demand. This is because they have some really old, dirty gas peakers that are on the margin during these high-demand hours. We finally get the plots of the sort you see in the figure, where for each gross generation level, we have damages per MWh.

Estimating marginal damage factors For every fossil fuel power plant we multiply hourly emissions of CO2 , SO2, NOX and PM2.5 by the county damages in $/ton in either AP2 or EASIUR (or the SCC for CO2) For each region (state, balancing area, eGrid, NERC), we run separate regressions by time of day and season: ∆damagesregion,h,season [$/h] = α∆generationregion,h,season [MWh] + ε  Results from the APEEP model give the average, dollar-per-ton damages for each pollutant (SO2, NOx, PM2.5, PM10, VOCs, and NH3) emitted by point sources. In APEEP, plants with an effective stack height higher than 500m are modeled individually for each facility, so the damages (in $/ton) from APEEP from those plants are used directly in our model. We use this assumption for 656 power plants. Plants with stack heights of less than 500m are modeled by APEEP as a county level emissions source (a point source in the middle of the county with either a low or medium effective stack height). For xyz power plants, we assumed that the plants have the average county-level point source damages (in $/ton of pollutant) estimates from APEEP for a specific county and for a specific stack height. For each source location, APEEP uses a Gaussian plume model to estimate the dispersion of emissions and the resulting concentrations in each county. Dose-response functions are used to estimate physical effects of affected populations and other receptors (crops, forests, materials, etc.). Physical effects are translated to monetary values using market prices for lost commodities, costs of illnesses, a value of a statistical life (VSL), and other non-market valuations from the literature. VOCs, NH3, and PM10 are excluded from this analysis because they result in damages that are, on average, more than two orders of magnitude lower than damages from other pollutants. At this point in the process we have hourly damages from individual power plants. The next step is to sum them up to find the houlry damages from all power plants in a region. And we do this separately for each of the 22 egrid subregions shown here. In this analysis we make the assumption that regions are completely isolated… we ignore imports and exports of electricity between regions. We then bin the data in 5% increments of gross generation. For each of these 20 gross generation bins, we regress the hourly change in damages per pollutant as function of the change in generations. We show a sample result for Texas. At high demand marginal CO2 and SO2 rates decrease as gas accounts for a larger share of marginal generation. In this case, the marginal NOx rates increase with demand. This is because they have some really old, dirty gas peakers that are on the margin during these high-demand hours. We finally get the plots of the sort you see in the figure, where for each gross generation level, we have damages per MWh.

The Marginal Factors tool is now available! We created a user interface with all our estimates being publicly available and downloadable for other modelers to use! https://cedm.shinyapps.io/MarginalFactors/

Average Marginal

Emissions Damages using AP2 Damages using EASIUR

Year Month Season and hour of day By load decile

NERC eGRID sub-region State

2010 to 2014 (now including 2015 and 2016)

SO2 NOx PM2.5 CO2

2014 time of day marginal emissions

I already received requests from several groups at LBNL, U I already received requests from several groups at LBNL, U. Minnesota, UC San Diego, Stanford U., etc, to use our estimates in their analysis Hopefully this tool will help with the wider use by modeler when performing policy evaluations We will continue to update the estimates, and also include a few other ways to measure marginal emissions