The effect of plug-in hybrid electric vehicle charging on fuel consumption and tail-pipe emissions

Plug-in hybrid electric vehicles can help reduce GHG emissions in the transport sector combined with the decarbonization of the electricity sector (EPRI 2007, Stephan and Sullivan 2008, Kromer et al 2009, Yang et al 2009, Poullikkas, 2015). Some studies focus on well-to-wheel GHG emission reductions based on fuel use and exclude GHG implications of vehicle manufacturing and disposal. Axsen et al (2011) use a survey data from new vehicle buyers in California and simulate the greenhouse gas emissions with one million new PHEVs on the road, and they conclude that PHEVs can cut marginal greenhouse gas emissions by one third to one fourth compared to conventional vehicles.

There are also studies that include GHG implications of vehicle manufacturing, battery production, and disposal that do a partial or full life cycle analysis (LCA). However, as Nordelöf et al (2014) points out most LCA studies regarding PHEVs lack a clearly stated and motivated goal and draw general conclusions without a proper discussion of the complexities of the outcomes. Shiau et al (2009) construct PHEV simulation models to analyse the impact of battery weight and charging patterns on GHG emissions; they conclude that as the lifetime monetary cost of a PHEV increases with battery size, the GHG emissions decrease. Michalek et al (2011) assess the economic value of PHEV life cycle air emissions and oil displacement benefits, and find that battery and electricity production emissions are substantial due in large to GHG emissions from coal-fired power plants, and conclude that PHEVs with small battery packs can reduce externality damages at a lower additional cost over their life time compared to PHEVs with long ranges. Plötz et al (2017a) analyse real-world fuel consumption data from 2,005 individual PHEVs of five different models and find that real-world direct CO₂ emissions decrease by 2% to 3% per every km of the all-electric-range. Plötz et al (2017a) also conclude that PHEVs charged from renewable electricity can substantially reduce well-to-wheel greenhouse gases, but electric ranges should not exceed 200 to 300 km because of the CO₂ intensity of battery production. Kamiya et al (2019) model short term and long term well-to-wheel effects of plug-in electric vehicles in Canada, and conclude PHEVs substantially contribute to reducing greenhouse gas emissions in all contexts explored. Plötz et al (2020a) analyze real-world usage and fuel consumption of approximately 100,000 vehicles in China, Europe and North America and find that for private vehicles real-world CO₂ emissions are two to four times higher than test cycle values. Wolfram and Hertwich (2021) perform a scenario analysis of different fueling behavior of PHEV users in the US—where they make varying assumptions on the level of mitigation challenges, electric carbon intensity, battery costs and carbon tax, all of which result in different levels of electric vehicle penetration—and conclude that the fueling behavior of PHEV users can determine a discharge of 0.7 GtCO₂ to 1.9 GtCO₂ over the next 30 years, emphasizing the importance of PHEV users' fueling and thereof charging behavior.

Existing studies in the literature focus on well-to-wheel GHG emission reduction but neglect the effect of charging behaviour on tail pipe emissions. Here, we fill this gap in the literature with the analysis on the effect of charging on tail-pipe emissions using real-world fuel consumption data with a long observation period for a long-range PHEV model in North America.

For our analysis, we use publicly available data representing real-world driving behaviour from an online source: Voltstats.net. Voltstats.net is an online database that collects real-world fuel consumption data of Chevrolet Volts, a plug-in hybrid electric vehicle, in the United States and Canada. Voltstats.net collects data automatically from an additional device. Voltstats.net was established as a volunteer effort by a Volt owner and a partnership with General Motor's OnStar in-vehicle safety and security system allowed users registered on Volstats.net to transfer their driving data automatically from OnStar. The data is automatically downloaded from the vehicles twice per day via OnStar and automatically transferred to the Voltstats website. The identification for the OnStar service is via the owner name and a vehicle identification number. The data contains daily values. Our sample consists of 7,491 reported Chevrolet Volt users with a user profile on the website containing daily data on the electric and gasoline mileage, including the number of gallons burnt per day by driving. Data was pre-processed, cleaned, and cumulative mileage values were converted to daily driven km. Daily VKT values larger than 1,500 km and observations of higher daily electric VKT than total daily VKT were excluded during the data cleaning process. Users with less than 28 driving days were also excluded from the analysis.

The dataset includes 3.4 million driving days with user specific performance data from April 2011 to January 2020. The average number of days observed per vehicle is 537 days with a median of 414 and maximum of 2,751 days; and average number of driving days per vehicle is 459 days with a median of 354 and maximum of 2,500 days.

From the cleaned data, the following parameters were calculated: electric VKT (eVKT), gasoline VKT (gVKT), total VKT, fuel consumption in litres of gasoline per 100 km and fuel consumption in charge sustaining mode in litres of gasoline per 100 km. The average daily values were extrapolated to annual values. The observed UF was calculated by dividing all electric km by total km driven during the observation period. The user specific profiles on Voltstats.net also include a statistic called EV-share, which corresponds to the UF calculated in our analysis. The EV-share statistic differs from UF per user on average ±2%, which we attribute to the automated calculation of EV-share that ignores anomalies such as unreasonably high daily VKT that we addressed already during data cleaning. Nevertheless, as a further step of precaution, we only included users with observed aggregated UF that were within ±10% of EV-share as stated on the user profile on Voltstats.net. In addition, three vehicles have been manually removed from the sample due to their unrealistic fuel consumption compared to their observed utility factor and daily driven km.

The user profiles on Voltstat.net do not provide information on the model year of the vehicle, so we use the base assumption that the date of the first logged trip for a vehicle indicates the model year of that vehicle. Based on this assumption, the following all-electric-ranges (AER) are used in our analysis: 56 km (35 US miles) for model years 2011–12, 61 km (38 US miles) for model years 2013–2015 and 85 km (53 US miles) for model years from 2016 onwards.

In order to estimate the frequency of additional charging and the frequency of no overnight charging, we compare calculated UF (UF_cal) and observed UF (UF_obs) for each day and user: UF_cal = AER/daily VKT if daily VKT > AER and 1 otherwise, and UF_obs = daily eVKT / daily VKT. Please note that UF_cal and UF_obs referred here are for each day per user and differ from the observed aggregated UF.

The calculation implicitly assumes a full recharge overnight. If the observed daily UF for a user is much higher than the calculated UF, the vehicle must have had at least one additional charge during the day. For the occurrence of such an additional charging event, we use the assumption that the observed UF for a vehicle for that given day is at least 1.5 times higher than the calculated UF. Similarly, for the occurrence of no overnight charging, we use the assumption that the observed UF for a vehicle for that given day is smaller than half the calculated UF. These assumptions are based on the observation of individual users which reveal that the ratio of observed UF to calculated UF for most users peaks at 1.0 (where UF_obs = UF_cal), and there is a valley around 0.5 and 1.5 where the ratio reaches low points. The frequency of additional charging is defined as the share of days with an additional charging event within the total number of driving days for a given user. Similarly, the frequency of no overnight charging is defined as the share of days with no overnight charging within the total number of driving days.

Our assumptions regarding additional charging and no overnight charging are rather conservative, which contributes to the robustness of our estimates. For instance, if a vehicle drives less than the AER on a given day and charges during the day, this occurrence will not be captured. Accordingly, some additional charging events during the day cannot be captured by our method and the obtained frequencies of additional charging are conservative estimates.

For the calculation of tail-pipe emissions, we use the values published by the U.S. Environmental Protection Agency, where on average 1 litre of gasoline (corresponds to 1 litre of fuel consumption in our dataset) produces approximately 2.31 kg of CO₂ (EPA 2005). After quantifying the effects of additional charging and no overnight charging on fuel consumption and tail-pipe emissions, we then perform a multivariable regression analysis to check for the statistical significance of their effect.

Table 1 shows the summary statistics of our dataset. The average observed aggregated UF is 76.6% with a median of 79.3%. Fuel consumption on average is 1.59 litres per 100 km with a median of 1.41 l/100 km and max of 6.35 l/100 km. We observe that the average share of days with additional charging is 9.1% with a median of 4.1%. The average share of days without overnight charging is 4.7% with a median of 2.7%. On average, users have a daily eVKT of 43.5 km, daily VKT of 59.0 km and annual VKT of 21,562 km. Please note that the annual mileage is close to the national average of 21,700 km in the U.S. (FHWA 2020).

We first look at the effect of not charging overnight on average fuel consumption. We find that regularly charging overnight — low frequencies of no overnight charging — reduces the mean fuel consumption below one litre per 100 km, see figure 1. Note that in figure 1 through figure 4, small dots represent users grouped and rounded to percentage values and blue line shows local average. We observe that higher frequencies of no overnight charging increase the mean fuel consumption and a share of above 60% nights without charging can push up the mean fuel consumption above 5 litres per 100 km. This significant difference shows that regularly charging overnight has a substantial effect on mean fuel consumption. The correlation between low charging frequency and higher vehicle emissions has a clear technical cause: If the battery has been fully recharged before the trip, then the battery will be fully depleted after the electric range has been exceeded. In that situation the combustion engine is used for propulsion of the vehicle and the battery can only buffer some energy from regenerative breaking. If the battery is not fully or only partly recharged before driving, the engine is needed for propulsion earlier or exclusively. Thus, low charging leads to more frequent use of the combustion engine and thus higher emissions.

Zoom In Zoom Out Reset image size

Given inset in figure 1, we control for daily VKT and look at the isolated effect of no overnight charging. This is done by looking at the difference between observed mean fuel consumption and calculated mean fuel consumption, where the calculated mean fuel consumption refers to 1- UF_cal multiplied by the fuel consumption in charge sustaining mode. From the inset in figure 1, we observe that regularly charging once overnight can result in a reduced observed mean fuel consumption of 1 litres per 100 km compared to a calculated mean fuel consumption, whereas not charging overnight 70% of the time can increase the observed mean fuel consumption by 3 litres per 100 km.

In figure 1, we observe that mean fuel consumption tends to increase in a steeper slope below 10% frequency of no overnight charging. We observe a different trend from 10% frequency of no overnight charging to 20% where the slope is less steep, and another trend with even a less steep slope when the frequency of no overnight charging is above 20%. We have run a piecewise linear regression for these three different trends and we find that the frequency of no overnight charging is statistically significant at the 0.1% level for all three trends. Here we provide the estimates and the standard error as added uncertainty (±): we find that fuel consumption and tail-pipe emissions increase by 1.85 ± 0.03 l/100 km or 42.7 ± 0.8 gCO₂ km⁻¹ tail-pipe emissions from 0% to 10% driving days without overnight charging (going from charging overnight everyday to only 9 out of 10 driving days). Around the mean frequency of no overnight charging (4.7%), mean fuel consumption is close to 2 litres per 100 km. We find that fuel consumption and tail-pipe emissions increase by 0.94 ± 0.12 l/100 km or 21.6 ± 2.87 g tail-pipe CO₂ per km from 10% to 20% driving days without overnight charging. Above 20% driving days without overnight charging, fuel consumption and tail-pipe emissions increase by approximately 0.42 ± 0.05 l/100 km or 9.73 ± 1.25 g tail-pipe CO₂ per km every 10% driving days without overnight charging.

The effect of regularly charging overnight on UF can be observed in figure 2. Looking at the difference between observed and calculated UF (main figure), we find that high shares of not charging overnight have a substantial effect on UF and not charging overnight 60% of the time can reduce the UF as much as 50 percentage points compared to the calculated UF that presumes charging every night.

Zoom In Zoom Out Reset image size

From figure 3, we observe that the effect of an additional charging event is less substantial compared to overnight charging, yet higher shares of additional charging results in lower mean fuel consumption. In figure 3, we observe that the mean fuel consumption is level around 1.6 l/100 km or 37 g tail-pipe CO₂ per km below 20% driving days with additional charging. A piecewise linear regression reveals that there is no statistically significant relationship between additional charging and fuel consumption when additional charging frequency is below 20%. Above 20% driving days with additional charging, regression analysis reveals a statistically significant relationship at the 0.1% level between additional charging and fuel consumption; and mean fuel consumption and tail pipe emissions decrease, on average by 0.08 ± 0.02 l/100 km or 1.86 ± 0.46 gCO₂ km⁻¹ tail-pipe CO₂ per km every 10% driving days with additional charging.

Zoom In Zoom Out Reset image size

Given inset in figure 3, if we control for daily VKT and look at the isolated effect of additional charging, we observe more clearly that an increase from 0% to 10% driving days with additional charging can result in a reduced observed mean fuel consumption of approximately 1 ¼ l/100 km or 29 g tail-pipe CO₂ per km. Above 10% driving days with additional charging, we find a reduced observed mean fuel consumption of approximately 0.3 l/100 km or 6.9 g tail-pipe CO₂ per km every 10% driving days with additional charging; e.g. 80% of driving days with additional charging can reduce observed mean fuel consumption by 3 l/100 km or 69 g tail-pipe CO₂ per km.

The effect of additional charging on UF can be observed in figure 4. Frequency of additional charging, as given in the inset of figure 4, has a smaller effect on observed UF compared to no overnight charging. However, if we control for daily VKT and look at the difference between observed UF and calculated UF, we see that additional charging around 80% of the time has the potential to increase the observed UF around 50 percentage points compared to the calculated UF.

Zoom In Zoom Out Reset image size

We use a multivariable regression analysis for a quantitative assessment of the effect of charging on fuel consumption. We distinguish between the frequencies of additional charging and the frequency of no overnight charging. Furthermore, we control for two additional variables with noteworthy impact on the UF and fuel consumption: the user's average daily VKT and the standard deviation (SD) of the daily VKT. The former indicates the typical daily driving distance while the latter also captures the variation in daily VKT where high SD is indicative of more frequent long-distance driving which additionally lowers the UF and increases fuel consumption at fixed mean daily VKT (Plötz et al 2018).

Our regression model is the following:

$$\begin{eqnarray}&&\begin{array}{l}FC={\beta }_{0}+{\beta }_{1}\,\mathrm{log}\left({f}_{{\rm{additional}}\,{\rm{charging}}}\right)+{\beta }_{2}\,\mathrm{log}\left({f}_{{\rm{no}}\,{\rm{charging}}}\right)\\ \,+{\beta }_{3}{\rm{Mean}}\,{\rm{daily}}\,{\rm{VKT}}+{\beta }_{4}{\rm{SD}}\,{\rm{daily}}\,{\rm{VKT}}+\varepsilon \end{array}\end{eqnarray} \tag{ 1 }$$

Where $FC$ denotes fuel consumption in litres per 100 km, f_{additional charging} is the frequency of additional charging (in %), f_{no charging} is the frequency of no overnight charging (in %) and the last two variables denote the mean and SD of daily VKT (both measured in km). We use the log of the charging frequencies as this reduces the likelihood of heteroscedasticity. See the appendix for the detailed discussion and robustness checks on heteroscedasticity and normality assumption in our model. The inclusion of the mean and SD of daily VKT reduces potential omitted variable bias.

The results of the regression analysis are given in table 2. Note that the number of users in the regression analysis is slightly less than what is presented in table 1 due to omission of users with 0% frequency of no overnight charging or additional charging.

The model itself and all variables are significant (mean daily VKT at 5% level and all others at 0.1% level) and have the expected sign. The simple model explains about 67% of the variance in fuel consumption, which is acceptable for the low number of variables included. An increase in no overnight charging, i.e. a decrease in charging leads to higher fuel consumption. Likewise, an increase in additional charging reduces fuel consumption and increases the UF. Higher mean daily VKT leads to a higher fuel consumption as a fewer share of km is driven on electricity. Finally, a higher SD of daily VKT indicates more frequent long-distance driving and thus lower UF coupled to higher fuel consumption.

We observe that both the log of frequency of no overnight charging and additional charging are statistically significant. For every relative 10% increase in the frequency of no overnight charging (e.g. from 10% to 11%), fuel consumption increases by 0.017 l/100 km (calculated by log(1.1)*0.42) and tail-pipe emissions increase by 0.40 gCO₂ km⁻¹. For every relative doubling of the frequency of no overnight charging (e.g. from 10% to 20%), fuel consumption increases by 0.13 l/100 km (calculated by log(2)*0.42) and tail-pipe emissions increase by 2.92 gCO₂ km⁻¹. On the other hand, for every relative 10% increase in the frequency of additional charging, (e.g. from 10% to 11%), fuel consumption decreases by 0.002 l/100 km and tail-pipe emissions decrease by 0.05 gCO₂ km⁻¹. Similarly, for every relative doubling of the frequency of additional charging, fuel consumption decreases by 0.015 l/100 km and tail-pipe emissions by 0.35 gCO₂ km⁻¹. For a concrete comparison, any relative doubling of the frequency of no overnight charging, for instance going from charging overnight 9 out of 10 nights (10%) to 8 out of 10 nights (20%) has almost ten times more impact on fuel consumption and tail-pipe emissions compared to a relative doubling in additional charging freqeuncy, for instance going from additional charging 4 out of 10 driving days to 8 out of 10 driving days. The regression analysis further establishes in a statistically significant way that no overnight charging has more impact on fuel consumption and tail-pipe emissions compared to additional charging.

Checking for multicollineratiy, the variance inflation factors for all variables in the regression model were low and less than 2.2. We tested the regression model with different treshold choices for the calculation of no overnight charging and additional charging frequencies. For the occurrence of additional charging, our assumption was that observed UF for a vehicle for that given day was at least 1.5 times higher than the calculated UF; we also tested for a threshold of 1.3 and 1.7. Similarly, for the occurrence of no overnight charging, our assumption was that observed UF is smaller than half (0.5) the calculated UF; we also tested for a threshold of 0.3 and 0.7. We only varied one threshold at a time, keeping the other same as in our base assumption. All cases of varying the threshold resulted in the same statistical significance for all variables and only slight differences in the estimation of the coefficients. We also tested the regression without the additional controls of mean and standard deviation of daily VKT and using the actual values for no overnight charging and additional charging frequency instead of the log of those frequencies. The results of these robustness checks on the regression analysis, including varying the thresholds, are given in table A1 in the appendix. We observe that the effect of frequency of no overnight charging is robust and significant in all cases, however the effect of frequency of additional charging is small and coefficient sign not robust without additional controls included in the regression. This is in line with our findings in sections 3.2 and 3.3. where we observe the significantly large impact of overnight charging on fuel consumption in figure 1, and that fuel consumption starts to decrease only after 20% of driving days with additional charging in figure 3. In our regression model, users with more than 20% additional charging frequency makes up approximately 14% of total users (886 users out of 6245). This shows that having high shares of days with additional charging (having more than 20% additional charging frequency) is limited to only a small percentage of all users and is not a common behavior; thus for the majority of users, additional charging has little impact on fuel consumption due to its low frequency.

In table A1, we provide the results of the robustness checks we performed on the regression analysis. Notice that the sample size varies under different assumptions regarding the threshold for the occurrence of charging events. This is due to the slight change in number of users with 0% frequency of no overnight charging or additional charging under each assumption, which are omitted from the regression analysis.

In table A2, we provide the robustness checks on heteroscedasticity. Our regression model uses the ordinary least squares (OLS) method which assumes that the error terms all have the same variance (homoscedasticity). We ran the Breusch–Pagan test to check for heteroscedasticity. Using the log of frequency variables reduces the Breush-Pagan test statistic from 523.9 to 349.2; however, the p-value is still less than 0.0001 which suggests that our model has heteroscedasticity. This shows that using the log of dependent variables reduces the likelihood of heteroscedasticity and makes our model more appropriate for OLS but it does not remove it completely. To estimate the severity of heteroscedasticity and whether it has a significant impact on our estimates and the significance of those estimates, we applied appropriate methods to deal with heteroscedasticity and compared it to our base model. The two most common ways to deal with heteroscedasticity is (1) to use heteroscedasticity-consistent standard errors and (2) to use the weighted least squares method (WLS). We provide the regressions results with both of these methods in table A2. Using heteroscedasticity-consistent standard errors does not change the coefficient estimates but aims to improve the standard errors by addressing the problem of errors not being independent and identically distributed. As shown in table A2, using heteroscedasticity-consistent standard errors produce the same result as our base model. Standard errors are slightly improved only in the 6th and 7th decimal places, which is not shown in the table due to rounding. All dependent variables also have the same level of significance. When we use the WLS method, we observe no visible changes in standard errors except for the intercept where it goes down from 0.06 to 0.05; there are slight improvements in the standard errors of our dependent variables but only in the 6th and 7th decimal places, therefore not visible in table A2. The coefficient estimates change slightly only for the intercept (from 1.79 to 1.60) and the log of frequency of no overnight charging (from 0.42 to 0.38). We also observe that the mean daily VKT loses its significance. Multiple and adjusted R-squared values are lower compared to our base model (from 0.669 to 0.650). Overall, WLS method does not provide any significant improvements, and it ends up resulting in a lower multiple and adjusted R-squared. In conclusion, these two methods (heteroscedasticity-consistent standard errors and WLS) provide no significant improvements on our model. This shows that the heteroscedasticity in our model is not severe and does not affect our coefficient estimates and standard errors in any significant way. Therefore, we kept our base model with OLS estimators.

We also test the normality assumption in OLS which assumes that the residuals are normally distributed. It should be noted that with larger sample sizes, the assumption of normality becomes less essential. This is due to Central Limit Theorem (CLT) which assures that the sampling distribution of the estimates will converge toward a normal distribution as N increases (Pek et al 2018). Statistical tests that are used to check for the normality assumption, such as the Anderson-Darling test and Shapiro-Wilks test, also become more difficult to interpret as the sample size increases. With large sample sizes, the power of the test increases such that it can find non-normal distributions with very small deviations. Given our large sample size, we used the graphical method of residual histogram and then checked for two statistical measures of shape, skewness and kurtosis to test for normality. The residual distribution of our regression model has a skewness of 0.29 and a kurtosis of 3.98. A skewness between −0.5 and 0.5 suggests that the distribution is approximately symmetric. Kurtosis for a normal distribution is 3; therefore, a kurtosis of 3.98 suggests that our residual distribution has a slightly higher central peak. The residual histogram is given below in figure 5. We observe no violation of the normality assumption.

Zoom In Zoom Out Reset image size