Electric vehicle charging is often celebrated as a clean alternative to gasoline, but a new study shows that when and how those vehicles plug in can determine whether the technology actually trims carbon emissions. Researchers report that a carbon-aware charging strategy, built on reinforcement learning and tested against real residential charging data, cut charging-related emissions by approximately 1.62% compared with uncontrolled charging, while delivering exactly the same amount of energy to the vehicles. The striking part of the finding is what did not work: an intelligent charging controller that optimized scheduling without any carbon information produced zero emission reduction, matching the uncontrolled baseline to the kilogram.
The study, published in the journal Carbon Research, reconstructed real-world charging sessions from an urban residential charging site at hourly resolution. Each session record contained a start time, an end time, and the total energy delivered. The team converted timestamps to a unified format, removed incomplete or invalid entries, and then spread each session’s energy across the hours it overlapped, preserving total energy so that no artificial loss or gain entered the analysis. This hourly demand profile was then paired with time-varying grid carbon intensity data from the Malaysian Peninsula power system, obtained from publicly available hourly emission records and expressed in kilograms of CO2 per kilowatt-hour.
The carbon intensity figures represent operational average emission intensity, calculated as total CO2 emissions divided by total electricity generation at each time step. The authors note that lifecycle emissions from infrastructure, fuel extraction, battery manufacturing, and vehicle production were excluded, since those components are independent of charging time and cannot be influenced by scheduling decisions. The analysis therefore captures the emissions that charging coordination can actually change.
Three charging strategies were compared under identical energy delivery and network constraints. The baseline allowed vehicles to charge immediately upon connection. A carbon-unaware variant of the Proximal Policy Optimization algorithm optimized charging subject to feeder capacity limits but ignored carbon intensity entirely. The proposed PPO-Carbon strategy explicitly incorporated grid carbon intensity into its reward function, penalizing emissions alongside feeder overload and charging delays. All three strategies delivered the same total energy of 53,740.38 kilowatt-hours over the evaluation horizon, ensuring that any difference in emissions arose purely from scheduling.
The results were unambiguous in their asymmetry. Uncontrolled charging produced 32,315.26 kilograms of CO2, with an average carbon intensity of 0.6013 kilograms per kilowatt-hour. The carbon-unaware PPO strategy produced identical figures, confirming that intelligent optimization alone does nothing for emissions when the objective ignores carbon. The carbon-aware strategy, by contrast, brought total emissions down to 31,793.24 kilograms and lowered the average intensity to 0.5916 kilograms per kilowatt-hour, a relative reduction of 1.62%, equivalent to roughly 522 kilograms of CO2 avoided.
The mechanism behind the reduction lies in a temporal mismatch the study quantified directly. Average grid carbon intensity during evening peak hours, defined as 18:00 to 22:00, measured 0.6789 kilograms of CO2 per kilowatt-hour, while off-peak intensity measured 0.5747 kilograms per kilowatt-hour, an 18.1% relative increase during peak periods. Residential charging commonly occurs in the evening when users return home, precisely when the grid relies more heavily on fossil-fuel generation. A representative month of data showed charging demand concentrated in the same hours when the grid was least clean, meaning a substantial share of uncontrolled charging happened at the worst possible times.
The carbon-aware controller did not simply delay charging uniformly. Hourly emission comparisons showed that reductions occurred predominantly during periods of elevated carbon intensity, particularly in the evenings, while emissions during cleaner hours closely tracked the baseline. Peak-period emissions fell from 9,330.80 kilograms under uncontrolled charging to 7,472.95 kilograms under the carbon-aware strategy, a pronounced drop, while off-peak emissions remained largely unchanged. This targeted shifting explains how the strategy achieved savings without curtailing energy or extending charging beyond users’ connection windows.
Statistical testing supported the finding. A Wilcoxon signed-rank test applied to paired daily aggregated emissions, chosen to reduce temporal autocorrelation in the hourly data, yielded a p-value of 0.002, indicating the reduction is significant at the 1% level. The team also checked robustness from several angles. A sensitivity analysis over the reward weighting coefficients showed that increasing the carbon weight monotonically reduced emissions, with diminishing returns once limited session flexibility was exhausted, and that the carbon-aware strategy consistently outperformed the baseline across realistic weight combinations. When Gaussian noise representing 5% and 10% forecast uncertainty was added to the carbon intensity signal, the reduction declined gradually but stayed above 1.4% and 1.2% respectively, showing the controller does not depend on perfect foresight.
How close does the learning approach come to the theoretical best? The researchers solved a deterministic linear programming benchmark with perfect knowledge of future carbon intensity under the same constraints. That optimum yielded 31,720.15 kilograms of CO2, only about 0.23% below what the PPO controller achieved. The small gap indicates the modest overall reduction is limited by the temporal flexibility of residential charging sessions and the actual variation in carbon intensity, not by algorithmic inefficiency. The authors also observed a trade-off: carbon-aware charging produced a slight increase in peak charging power, because the objective prioritizes emission reduction over peak load minimization and can concentrate charging in low-carbon intervals that are not the lowest-load periods.
The authors caution that the analysis covers a single urban residential feeder, which constrains the magnitude of achievable reductions, and that historical carbon intensity data may shift as renewable generation grows. Feeder constraints were modeled as an aggregate capacity limit, without detailed voltage regulation or power flow modeling, and electricity cost optimization was not considered. Even so, the study demonstrates a clear principle: emission savings from electric mobility depend on aligning charging with cleaner electricity hours, and that alignment requires carbon signals to be built directly into the control objective. Without them, even sophisticated learning-based scheduling leaves the carbon footprint untouched.
Subject of Research: Quantifying carbon emission reduction from carbon-aware electric vehicle charging in distribution power networks
Article Title: Quantifying carbon emission reduction from carbon-aware electric vehicle charging in distribution power networks
Article References: Hossen, M. S., Sarker, M. T., Ramasamy, G., & Eng, N. E. (2026). Quantifying carbon emission reduction from carbon-aware electric vehicle charging in distribution power networks. Carbon Research, 5(1), Article 43. https://doi.org/10.1007/s44246-026-00280-6
Image Credits: AI Generated
DOI: 10.1007/s44246-026-00280-6
Keywords: electric vehicles, carbon-aware charging, reinforcement learning, grid carbon intensity, distribution networks, emission reduction, smart charging, PPO, residential charging, load shifting, decarbonization, Malaysia
News Source: Bethany Barker. (November 5, 2026). Carbon-aware EV charging cuts emissions 1.62% without extra infrastructure. Scienmag.
