Air quality trade-offs of a rapid expansion of personal electric vehicles in China


 In China, replacing gasoline cars with electric vehicles (EVs) is at the center of a strategy to reduce air pollution and CO2 emissions from transportation. Previous estimates of the benefits of vehicle electrification quantified the impact of EV use on on-road and power generation emissions only, thereby neglecting gasoline production. This study presents the first “use-cycle” analysis of EVs in China, including changes in emissions from transportation, power generation, and oil refineries. We use the GEOS-Chem atmospheric chemistry transport model to quantify how each sector contributes to the net impacts of EV use on air pollution (PM2.5 and ozone) in China. We find that the projected growth in EV usage by the end of 2020 results in ~1,900 (95% CI: 1,600–2,200) avoided premature mortalities annually and a 2.4 Mton decrease in CO2 emissions. 70% of the total reduction in mortality is due to avoided refinery emissions. As refinery emissions become more tightly regulated, our work implies that the power generation sector must also become cleaner for EVs to remain beneficial.


Introduction 1
Outdoor air pollution in China causes over 900,000 premature mortalities each year 1,2 , and its 2 reduction is a regulatory priority 3 for the country. In recent years, sector-specific policy strategies 3 have been successful at reducing the emissions of air pollutants and their precursors [4][5][6][7][8] . In the 4 transportation sector, EVs are considered a key element of the strategy to reduce air pollution-5 related health impacts. In 2015, an estimated ~25,000 premature mortalities in 2015 12 (~3% of 6 the annual total from all sources 13 ) were attributed to road emissions and ~49,000 12 (~6% of the 7 annual total 13 ) to the fuel processing sector. 8 9 The transition to EVs will also result in increased demand for electric power generation due to 10 vehicle charging. Power generation emissions are already one of the largest contributors to 11 outdoor air pollution in China, causing ~100,000 premature mortalities in 2015 (95% CI: 50,000-12 180,000) (11% of the yearly total) 9,11,12 . When evaluating the net impacts of the use of EVs on 13 air pollution and CO2 emissions, it is therefore necessary to quantify its effects on all three of 14 these systems: transportation, fuel processing, and power generation. This study is the first to 15 present such a vehicle "use-cycle" analysis to answer the question of whether replacing 16 gasoline cars with EVs in China results in net air quality benefits. 17 changes in air quality (considering PM2.5 and ozone). The health impacts of these changes are 1 computed using concentration-response functions derived from the epidemiological 2 literature 31, 32 . 3 4 The impacts of EV deployment are calculated with respect to a reference case with no 5 additional EVs on the road in 2020 compared to 2017, so that the difference in energy demand 6 from passenger cars between 2020 and 2017 is met by new gasoline cars. Emissions from 7 transportation and refineries in the reference case assume a fuel efficiency of new gasoline cars 8 of 5 L/100 km, corresponding to the 2020 fuel-efficiency standard. This provides our baseline 9 estimate for air quality. Our EV scenarios are based on an estimate of EV penetration by 2020 10 in each province from the ECLIPSE dataset 33 and assume that EVs substitute 1.   (Table 1). Compared to a "slow charging" case, fast charging has been shown to 20 increase peak electricity demand and result in a disproportionate allocation of the EV load to 21 fossil-fuel sources, thereby increasing total emissions 27 . Least-cost smart charging ("smart 22 charging" hereafter) can reduce generation costs by mitigating demand peaks but may not 23 reduce emissions. Finally, we simulate "e-smart charging", in which carbon pricing is applied 24 only to the EV charging load, thereby incentivizing the use of renewable electricity sources for 25 EV charging. Because the emissions response to the power demand from EVs changes with the 1 power grid, we simulate it with the 2017 power grid, and the projected 2022 power grid.

5
These four scenarios represent limiting cases, but serve to provide bounds to the real-world 6   In both the slow and fast charging scenarios, electric demand for EVs peaks during the early 6 evenings. It coincides with the pre-existing daily peak in total electricity demand, thus resulting 7 in relatively high emissions. Smart charging assumes that EV charging can take place at any 8 point in a predefined 12-hour window, so that the demand can be fulfilled by cheaper, 9 underused gas and hydropower outside of peak hours. As such, smart-charging can reduce 10 emission intensities, although emissions intensities are still 41-74% greater than the grid 11 average. Smart charging with a carbon price for the EV load ("e-smart charging") further 12 reduces NOx and SO2 emissions to 54% and 95% lower than the power grid average emission 13 factors, but PM2.5 and CO2 emissions intensities are still 55 and 37% higher than the grid 14 average. This indicates that while e-smart charging does not completely divert the EV load from 15 coal to renewable sources, it dispatches the EV load to cleaner coal-fired power plants than 16 smart charging. 17 18 The spatial distribution of the increase in power generation NOx emissions in the slow charging 19 scenario is shown in Figure 2  In some provinces, total emissions are reduced by the national expansion of EVs (Figure 2 (e) 1 and (f)). NOx, SO2, PM2.5, and CO2 emissions decrease in Beijing (0.2%, 0.4%, 0.1%, and 0.3% 2 respectively), but this is accompanied by increases from the nearby Hebei, Inner Mongolia, and 3 Shanxi provinces. In total, 20 of the 29 provinces under consideration show increases in one or 4 more of NOx, SO2, PM2.5, or CO2 emissions, while 22 provinces show decreases in one or more 5 of these pollutants.   The spatial heterogeneity of emissions changes is due to the heterogeneity of demand, but also 1 due to some provinces importing a large fraction of the electricity required to charge EVs. For 2 example, Beijing and Tianjin are net importers of "EV electricity": we estimate that they generate 3 less than 10% of the electricity demand for EVs locally. Provinces such as Shanxi, Xinjiang, and 4 Hebei are instead net exporters. The increase in power production in Shanxi, for example, 5 corresponds to 1.7 times the increase in local demand. This reflects the growing reliance of the  We propagate these emissions changes through to annual premature mortalities using 6 concentration-response function (details are in the Methods section). The total number of 7 premature mortalities estimated for each scenario relative to the baseline is summarized in 8    with outdoor air pollution in China 37 ). Fast charging results in the smallest improvements in air 12 quality compared to the baseline (~1,800 (95% CI: 1,500 to 2,100) avoided premature 13 mortalities), while smart charging yields 26% greater air quality benefits than slow charging 14 compared to the baseline). Smart charging with carbon pricing ("e-smart charging") results in 1 the largest net air quality benefits relative to the baseline case (~3,800 (95% CI: 3,400-4,100) 2 avoided premature deaths). This is because of a cleaner power grid response. This scenario 3 also has the largest province-level benefits, with annual premature mortalities in Beijing 4 decreasing by 140 (95% CI: 90-180). We also find 200 (95% CI: 160-240) prevented deaths in 5 Chongqing and 70 (95% CI: 50-90) in Tianjin. benefits under all four EV scenarios. Figure 5 breaks down the net health impacts into the 20 contribution attributable to changes in emissions from power plants, on-road, and refineries 21 emissions. In the slow charging scenario, we find that reductions in emissions from refineries 22 are the main driver of reduced exposure to ozone, contributing 93% of the total reduction in 23 premature mortalities from ozone exposure and 69% of the reduction in PM2.5 exposure.   The heterogeneity of net impacts is further illustrated for the slow charging scenario in Figure 7. 5 In Central and Northeastern China, EV penetration leads to a net decrease in mortality. In 6 Southern China, air pollution-related mortality increases with EV penetration. Most of the 7 country's major cities are in Northeastern and Eastern China, and the concentration of 8 population in these regions amplifies the benefits of reducing on-road emissions, while recent 9 efforts to develop long-distance transmissions have reduced the impact of power generation on 10 these regions' air quality 36 . Northeastern China also concentrates most of the oil refining 11 capacity, and therefore captures most of the benefits of reductions in refineries emissions (see 1 Supplementary Information). Overall, reductions in air quality-related premature mortalities and 2 electric demand for EVs in each province are only partially correlated (r 2 =0.17). This reflects the 3 fact that the air quality of a given province is influenced by emissions outside its borders, and 4 suggests that the introduction of EVs at the province-level alone might not be sufficient to 5 reduce air pollution locally while it may negatively impact neighboring provinces.

12
These results are also sensitive to policy choices regarding vehicle availability and generation.
induced demand could be driven by changing consumer behavior for EVs, e.g., in response to a 1 lifting of existing vehicle licensing restrictions for EVs in cities such as Beijing and Shanghai, or 2 to lower cost of ownership which would make EVs more affordable than gasoline cars. With 3 substantial induced demand, the air quality benefits of EVs could be reduced or even reversed. 4 We find that, if EV deployment does not displace gasoline vehicles and the reductions in 5 refinery and on-road emissions are not included, then EV deployment would translate into 6 ~2,700 (95% CI: 1,800-3,500) additional premature mortalities and 23 Mton of CO2 emissions in 7 the slow charging scenario. 8 9 Our study suggests that the deployment of EVs in 2020 already has regional benefits to air 10 quality in Central and Northeastern provinces compared to a baseline case with gasoline cars. 11 However, net air quality benefits at the national scale could more than double if renewable 12 sources were deployed to match the increase in electricity demand. 13 14

Sensitivity analysis and discussion 15
To account for uncertainties in the emission factors of power plants as reported in previous 16 studies 42,43 , we repeat our analysis for the slow charging scenario with a uniform, 50% reduction 17 in NOx emissions from power plants. The net benefit of the EV deployment on air quality in this 18 case further increases by 42% (accounting for reduced on-road and refineries emissions). A 19 similar scenario with a 50% reduction in SO2 emissions increases the air quality benefits of the 20 EV deployment by 11%. 21

22
In addition, we acknowledge that energy policies can have profound impacts on the outcomes. If 23 installation of renewable capacity was increased such that new renewable sources exactly 24 matched the EV load, the negative outcomes associated with increased power plant emissions 25 could be avoided, and EVs would result in ~4,600 (95% CI: 3,700-5,400) avoided premature mortalities due to improvements in air quality. Such a scenario would also imply a 26 Mton 1 reduction in annual CO2 emissions. These reductions may be underestimated, since we assume 2 that the displaced gasoline cars in our baseline case meet the 2020 fuel economy standard of 3 5L/100 km (compared to an average fuel economy of 5.8 L/100km in 2018 41 ). 4

5
Most studies on the impacts of EV deployment on emissions recommend the use of marginal 6 emission rates from power plants 27,46-49 , consistent with this study's consequential approach. 7 However, others argued that this approach is not well-suited for long-term studies 50 because 8 new electricity generating units are built over time in response to projected increases in 9 demand. To quantify the sensitivity of our results to the load allocation approach, we develop a 10 sensitivity case where the EV load is allocated exclusively to recent generating capacity instead 11 of consequential capacity (i.e. capacity built after 2017 and equipped with the most stringent 12 emissions controls; see Methods). In this case, the modeled EV deployment results in ~3,800 13 (95% CI: 3,100-4,500) avoided mortalities nationwide and a reduction of 13 Mton in CO2 14 emissions. In addition, EV deployment is beneficial not only in Northeastern and Central 15 provinces but also in Southern provinces (see Supplementary Information). This reinforces the 16 fact that if the EV load in these provinces were matched with cleaner generation, the benefits 17 associated with EV deployment would increase significantly. 18

19
It has previously been reported that the GEOS-Chem model may overestimate the production of 20 nitrate aerosol. We quantify the sensitivity of our results to nitrate aerosol by performing an 21 additional set of simulations in which 25% of HNO3 is removed at every time step 44,45 . This leads 22 to a 27% increase in estimated net benefits of slow-charging EVs. In these cases, the spatial 23 distribution of net air quality impacts is unchanged and EV deployment is more beneficial in 24 Central and Northeastern China than in Eastern and Southern China. Results obtained with premature mortalities to ~2,300 (95% CI: 1,700-2,900) instead of ~1,900 (95% CI: 1,600-2,200) 1 (detailed results from the sensitivity analyses are presented in the Supplementary Information). 2 3 Additionally, the importance of a cleaner power grid will likely increase when recently 4 announced restrictions on air pollutants' emissions from oil refining facilities 28,30,51 are fully 5 implemented. As emissions from refineries decrease, so will the air quality benefits associated 6 with displaced gasoline consumption from the deployment of EVs, which currently make up 70% 7 of the total air quality benefits of EVs. Absent the benefits from reduced refineries emissions, 8 the dominant driver of air quality changes under the EV scenarios is power plant emissions 9 ( Figure 7), which result in increases in ambient PM2.5.Therefore, to ensure that EVs continue to 10 contribute to a reduction in air pollution, stricter emissions controls on power plants and further 11 investment is renewable energy are needed. 12

13
Our estimate for the CO2 reductions due to the EV deployment, however, does not account for 14 emissions associated with the production of EVs, in particular their battery. Battery 15 manufacturing alone has been shown to make electric vehicle production 30% more GHG-16 intensive than gasoline vehicle production 53 , and including them would reduce the calculated 17 benefit of EV deployment on CO2 emissions. The impacts of battery manufacturing on air quality 18 has not been explored in detail. 19 20 Other benefits of EVs in China not accounted for are the reduced reliance on petroleum 21 products imports, support for domestic manufacturing in the automotive sector, or mitigation of 22 the urban heat island effect 54 . These factors can be considered in future cost-benefit analyses of 23 EVs at the national level. This work does not account for potential changes in annual mileage 24 driven when replacing gasoline cars with EVs. Finally, taking into account the fact that no single charging mode is used for all EVs would better reflect actual on-road behavior and provide more 1 granular insights on the real-world effect of EVs on air quality. Additionally, the air quality and climate benefits associated with EV deployment are increasing. 21 If we assume a cleaner power grid (consistent with 2022 projections), the net reductions in mortality due to EVs would be increased by 26% in the slow charging scenario, while net CO2 23 emissions would further reduce by 4% compared to 2017. Recently announced emissions 24 controls on oil refining facilities would reduce the benefits of EV deployment on air quality and 25 CO2 emissions, unless cleaner power generation is deployed to match the EV load.

Methods 1
The following sections describe the methods used to obtain the results presented in the Main 2 text. Section 5.1 presents the power grid model. Section 5.2 details how the demand for EV 3 charging is estimated and allocated. The method to estimate reductions in on-road and 4 refineries emissions is presented in Sections 5.3 and 5.4. Section 5.5 describes the air quality 5 model, and Section 5.6 details this study's approach to health impacts analysis. Under the smart charging and e-smart charging scenarios, the EV demand at any given hour is 11 not required to be met within the same hour, and the grid operator may allocate the EV load 12 during any of the following 12 hours. In the smart charging scenario, the allocation within that

Electric demand for EVs 15
This section details how the EV demand was derived and allocated to specific hours of the 16 years. Section 5.2.1 presents the derivation of the total energy used for personal vehicle 17 transportation in the EV and counterfactual scenarios. Section 5.2.2 outlines how the electric 18 demand for EVs was allocated to each hour of the year. 19 20

Estimating total annual demand by province 21
The annual power demand from EVs by province is taken from the GAINS China model 22 (ECLIPSE CLE scenario v5a 33 ) and totals 88.49 PJ for 2020. Projections for the EV demand are 23 derived from the objectives formulated in the 12 th Five-Year Plan that covers the period 2011-responsible for 3.5% of distance traveled in 2020 under this projection. They will also be 1 responsible for 0.38% of the total grid electrical energy demand in China, assuming no change 2 from 2017. The projected energy demand for EVs by province is shown in Figure 8.

Temporal profile of the EV demand for each scenario 9
To allocate annual EV demand to individual hours of the year, we follow the approach 10 developed in Chen et al. 27 . The approach distinguishes demand patterns for weekdays from 11 demand patterns for weekends, and considers EV charging at home, at work, or at shopping 12 centers. In each category, each day is assumed to bear the same share of the annual EV load. 13 Hourly demand for EV charging on a typical weekday is represented in Figure 9 for the slow and 14 fast charging scenarios. During the week, three charging locations are considered: workplace 15 (7kW, in the mornings, making up 20% of the daily EV load), home (3kW, at night, representing 16 70% of the daily EV load) and shopping centers (7kW, at night, representing 10% of the daily and the other half at home. In the fast charging case, the charging rate at all locations is set at 1 60 kW, a typical value for fast charging 27 . 2 3 Despite the fact that all provinces operate under the same time zone, the timing of electricity 4 demand varies between provinces 27 . To account for this profile, EV load is allocated to 5 individual hours based on the apparent time zone at each of the provinces' centroid. This 6 approach preserves the ratio of EV demand to general electric demand throughout the day 27 . 7 The time difference between UTC time and the apparent time zone is defined as the longitude 8 of the province centroid divided by 15 and rounded to the nearest integer. Finally, in all 9 scenarios, we consider that the electricity required to charge one EV is 12.8 kWh (the average 10 value used by Chen et al. 27 ), and that each EV is fully charged once per day. 11 12 Figure 9. Additional hourly demand due to EVs in Beijing in the slow charging scenario (blue) and fast charging 13 scenario (red) on a weekday.

15
The power dispatch model used in this study does not consider transmission constraints within 16 provinces, so that the EV demand by province is not further spatially disaggregated. In the 17 smart and e-smart charging cases, electric demand for EVs is input into the power grid model as a constraint over 12-hour windows instead of 1-hour windows. This allows the EV load to be 1 dispatched at any time within 12 hours of the first connection to the grid at the lowest cost.

Emissions from refineries 23
Changes in refineries emissions are derived from the total gasoline consumption avoided under 24 the EV scenarios obtained with the method described in Section 5.3. and refineries emissions 25 intensities from Zheng et al. 29 . Following recent data on imports of oil products, we assume that 95% of the gasoline consumed in China is refined in China; the remaining 5% is imported 35 . The For ozone, the results reported in the main text are obtained using a log-linear concentration-1 response function with parameters derived from Turner et al. 32 . Turner et al. associated 2 exposure to 8-hour maximum ozone concentration to premature mortality from respiratory and 3 circulatory diseases (ICD-10 codes I00-I99 and J00-J99) using a two-pollutant model adjusted 4 for PM2.5. They found a central relative risk for circulatory diseases of 1.03 (95% CI: 1.01 to 5 1.05) and a central relative risk of 1.12 (95% CI: 1.08 to 1.16) for respiratory diseases. As for 6 PM2.5, we generate 10,000 samples for each of these parameters to estimate ozone-related 7 health impacts. The form of the log-linear function relating relative risk to the annual mean 8- The relative risk for each endpoint and age group is related to the number of premature 15 mortalities as 16 3,9 = 9 × 3,9 × RR 3,9 − 1 RR 3,9 17 where 3,9 is the number of premature mortalities from endpoint and age group in a given 18 grid cell, 9 the population in the age-group in that grid cell, 3,9 the baseline incidence, and 19 RR -,: the relative risk obtained using the CRFs described above 76 . The following sections present the spatial distribution of population in China in Section S1, 3 additional results with the 2017 power grid in Section S2, detailed results with the 2022 power 4 grid in Section S3, additional scenarios run as sensitivity analysis in Section S4, and validation 5 data for the power grid model and for the air quality model in Section S5.        To identify the effect of changes in the Chinese power grid, we repeat the power grid analysis 6 with the projected 2022 power grid. As older power plants are phased out and new, less coal-7 reliant capacity with state-of-the-art emissions controls is installed (see Methods for details), we 8 find that the consequential capacity that is dispatched in the EV scenarios is also less reliant on 9 coal and has lower emission factors than its 2017 counterpart. Similarly to the 2017 power grid, 10 we estimate that the consequential capacity available to meet the EV load under the slow, fast, 11 and smart charging scenarios still has higher emission factors than the grid averages (54-86%, 12 57-95%, and 55-71%, respectively). Figure S4 below summarizes these findings.

4
We estimate that premature mortalities under the slow and fast charging scenario reduce in 5 Eastern and Southern provinces, where we predict that the 2022 power grid will allow for net air 6 quality benefits in these provinces. On the contrary, Northeastern provinces will see an increase 7 in premature mortalities under these two scenarios. Smart charging with carbon pricing (e-8 smart) with the 2022 power grid had smaller air quality benefits than with the 2017 power grid In 9 particular, Northeastern provinces experience increased mortalities under the e-smart charging 10 scenario. However, the total number of premature mortalities decreases under all scenarios 11 compared to the 2017 power grid case, as summarized in Figure 7 of the main text and Figure  12

S4.2. Result sensitivity to the EV load allocation
Despite the fact that no major grid changes occur during the EV deployment considered in this 1 study 38 , some new generators are deployed along with EVs between 2017 and 2020. To 2 account for the fact that under such circumstances, the consequential approach used in the 3 main text does not fully capture the emissions change, we consider a sensitivity scenario where 4 the EV load is matched exclusively by power plants built after 2017 and equipped with the most 5 stringent emission control devices (see Methods for details). In this case, we find that the 6 modeled EV deployment under the slow charging case with the current becomes net beneficial, 7 reducing premature mortalities by 160 nationwide and CO2 emissions by 5.6 Tg. In this 8 scenario, the spatial distribution of impacts is also modified (see Figure S9). 9 10 11 Figure S9. Air quality impacts associated with the modeled EV development in the slow charging conditions 12 in the slow charging scenario described in the main text (left) and assuming that the EV load is matched 13 exclusively by recent generators (right). 14 15 In this sensitivity case, 14 out of 29 provinces have reductions in air pollution-related mortalities. 16 The major difference is the Northeastern provinces. In the main case, EV deployment is 17 associated with an increase in mortality, while in the sensitivity case they have net air quality 18 benefits associated with EV deployment. This further highlights the point made in the main text that the deployment of a cleaner grid in the Northeast has the potential to avoid the negative 1 impacts calculated. to changes in PM2.5 concentrations in the slow charging scenario decreases by 65% to ~300 7 (95\% CI: 100-500) avoided premature mortalities. With a central estimate for ozone impacts of 8 ~1,100 (95% CI: 800-1,400) fewer premature mortalities, this means that the net impact of the 9 slow charging scenario is ~1,400 (95% CI: 800-2,000) avoided premature mortalities per year. 10 The spatial distribution of impacts is not modified. In contrast, when using the CRF from Hoek et 11 al. 75 , the number of PM2.5-related mortalities increases by 13% to ~900 (95% CI: 700-1,100) 12 avoided premature mortalities, bringing the net number of avoided mortalities to ~2,000 (95% 13 CI: 1,500-2,500). 14 15 In the case of ozone, using the CRF based on the daily 1-hour maximum during the ozone 16 season from Jerrett et al. 83 yields 33% greater reduction in ozone-related premature mortalities, 17 bringing the total number of avoided mortalities in the slow charging scenario to ~2,300 (95% 18 CI: 1,700-2,900). The spatial distribution of the impacts is not modified. 19 20 S4.4. Result sensitivity to the GEOS-Chem nitrate mechanism 21 production 45 . One proposed solution is to decrease the HNO3 concentration in the input to the 23 thermodynamic gas and particle partitioning by 25% at each time step 44 . We implement this fix 24 and find that the total number of premature mortalities in the slow charging scenario decreases 25 by 36% to -2,200. It reduces the impact of EVs in all scenarios on PM2.5 concentrations by 32%, but also reduces the total benefits of reducing on-road and refineries emissions. These results 1 should however be interpreted with caution as the corrected mechanism has been questioned 45 2 and the corrected simulations yield a lower r 2 -correlation with monitor data (see Section S5.2). 3 4 S5. Validation of the power grid and air quality models 5 This section details the methods and the results obtained when comparing the results in our 6 power grid (S5.1) and air quality (S5.2) models to data from the literature.