Meet SynthCharge: Smarter Synthetic Data for EV Charging
The “average EV driver” doesn’t exist. Some drivers plug in every evening; others only charge when their battery is nearly empty. Some remain connected overnight, while others need a rapid top-up before their next journey. Some drive only a few thousand miles a year,while others regularly make long-distance journeys.
These differences matter when designing smart-charging services, electricity tariffs, and customer propositions. Yet EV tariffs and smart-charging propositions, including the savings communicated to customers, are often assessed using a simplified representation of the “average” driver.
Building more representative propositions requires detailed data on how individual customers charge. But real-world charging datasets are difficult to access and share because they can contain sensitive information about household routines and driving needs. Public datasets help, but they are often geographically fragmented or too limited to represent the diversity of charging behaviour at scale.
Example EV charging sessions generated by SynthCharge over one month for 200 EV driver profiles across two groups: Battery Electric Vehicles on a time-of-use tariff and Plug-in Hybrid Electric Vehicles on a flat standard variable tariff.
To deal with this challenge, Energised Futures has developed SynthCharge, a generative AI model that creates realistic synthetic EV charging sessions.
Rather than just modelling the average driver, SynthCharge is able to model populations of EV customers using relationships learned from the charging behaviour of thousands of customers, including when vehicles are plugged in, how long they remain connected, and the amount of energy delivered during each charging session. SynthCharge can generate new charging data conditioned on characteristics such as annual mileage, tariff type, and battery capacity. This allows users to explore different customer groups and scenarios, rather than relying on a single representative driver, helping to reveal how outcomes vary across different travel needs, charging routines and levels of flexibility.
This gives researchers and proposition teams a practical way to explore a wider range of EV charging behaviours without needing direct access to sensitive customer records. By generating realistic populations of charging sessions, SynthCharge can support tariff testing, grid-impact studies and analysis of how different customers could benefit from smart charging.
How it works
An EV charging session tells us both when a vehicle is available and the amount of energy delivered. These characteristics are closely related: the time of arrival and the length of the connection affect how much energy can be delivered. Charging behaviour also follows daily, weekly and seasonal patterns, all of which the model needs to reproduce.
When developing SynthCharge, we looked at conventional approaches first, which describe each charging session as a separate event, recording attributes such as arrival time, connection duration and delivered energy. In our experiments, this event-based representation struggled to reproduce recurring charging patterns over the given time period. Instead, we converted each vehicle’s charging history into two timelines, divided into 30-minute intervals (see Figure below). One timeline shows when the vehicle is plugged in, the other shows how much energy is delivered during those periods. This does not recreate the vehicle’s exact charging power at every moment, but it gives the model a clear picture of when and how much people charge. Of the approaches we tested, this produced the most accurate results and allowed the model to more consistently recognise repeated charging habits over days and weeks.
SynthCharge then uses a conditional denoising diffusion probabilistic model, or CDDPM, to learn the patterns contained in these paired signals. The generation process is conditioned on customer, vehicle and infrastructure characteristics such asestimated annual mileage, tariff type, battery capacity and vehicle class. This enables SynthCharge to generate populations of charging sessions that are consistent with specified characteristics, rather than producing a single generic representation of EV charging behaviour.
Use cases
SynthCharge enables scenario-driven analysis of grid conditions, tariffs and EV charging behaviour. To demonstrate this, we developed an interactive dashboard that allows users to define a population using characteristics such asannual mileage, vehicle class, tariff and battery capacity.
SynthCharge then generates charging sessions consistent with those conditions, allowing different tariff designs to be tested across customer groups and scenarios. Rather thanreporting savings for a single representative driver, the dashboard reveals how outcomes vary across different travel needs, charging routines and levels of flexibility. This can help us to identify where savings are robust, where theydepend strongly on customer behaviour, and where tariff communication may need to be more transparent.
Performance and benchmarking
We evaluated the proposed generative architectures against five criteria:
- Marginal distribution fidelity: Do generated - charging sessions and energy demands match real -data?
- Joint dependency preservation: Does the model reproduce the relationship between connection duration and delivered energy?
- Temporal-cycle accuracy: Does it reproduce time-of-day, day-of-week and other periodic behavioural patterns?
- Physical-constraint adherence: Do generated charging sessions stay within what is physically possible?
- Downstream predictive utility: Can models trained on synthetic data perform effectively when evaluated against real charging?
Across the 5 criteria, SynthCharge achieved the strongest overall balance between distributional fidelity, dependency preservation and downstream utility, although no model performed best on every metric. The model does not yet enforce all physicalconstraints, and some generated sessions exceed the applicable maximum-energy condition.
For comprehensive benchmark results, read our full paper
Conclusion
SynthCharge demonstrates that conditional diffusion models can generate EV charging sessions that retain important temporal, behavioural and physical relationships from real-world data. Our results show the potential for synthetic data to provide a more representative picture of EV charging behaviour, helping us move beyond assumptions based on the “average” driver.
There are still clear areas for further development, particularly around stronger enforcement of physical constraints and validation across countries, charging environments and vehicle types. But SynthCharge provides a strong foundation for exploring how different customers charge, and what that could mean for future tariffs, smart-charging services and the electricity system.
If you are interested in potential collaboration on this subject, you can contact us at [email protected].
