They Cloned a Decade of EV Charging to Protect a Norwegian City's Grid

Somewhere in Trondheim, Norway, on a cold February morning in 2030, a row of electric cars will sit plugged into chargers outside apartment blocks. The grid operator will need to know, to the hour, how much power that street is about to draw — because if everyone plugs in at once after a cold commute home, the local transformer could overheat, brown out, or fail outright. That question isn't hypothetical. Norway already has the world's highest electric-vehicle adoption rate, and its distribution grids are the test bed for a transition the rest of the world is watching.

The problem is that we can't actually see that 2030 morning. Nobody knows its weather, its holiday calendar, or how many new EV owners will have arrived by then. So a team of researchers built it — not by guessing, but by teaching a machine to imitate 14 months of real Trondheim charging logs, then projecting that behavior forward through an entire decade. The result is a synthetic dataset of 76,993 hourly charging records spanning February 2020 to December 2030, representing 256,826 individual charging sessions and roughly 3.26 GWh of delivered energy (Taraghi Nazloo & Musilek, 2026).

This isn't a paper about what EVs do to the grid today. It's about equipping planners to handle what happens when the cold, dark winters of a Scandinavian city meet a fleet that keeps growing. And the single most bracing number buried in it is this: a peak hourly aggregated demand of 1,559.87 kWh — a load spike that, if unmanaged, would strain any distribution transformer it lands on.

The Science

Let's talk about what a "synthetic EV charging dataset" actually is, because the phrase hides something clever. Real charging data is precious but limited: it covers one city, one set of infrastructure, one narrow window of time. Trondheim's observed logs run from December 2018 to January 2020 — a mere 14 months. You can't plan a decade of grid investment on 14 months of history, especially when the fleet itself is expanding year after year.

So the researchers (Taraghi Nazloo and Petr Musilek, based in Edmonton, Canada) built a generative pipeline that learns the pattern of real charging behavior and then re-samples it into the future. Think of it less as a crystal ball and more as a very sophisticated statistical cloning machine.

The architecture has several layers. First, they cleaned the historical session logs, removing records with missing timestamps, negative energy values, or inconsistent data — while deliberately keeping long-duration and high-power sessions that might look like anomalies but actually represent authentic behavior (a car plugged in for days at an airport lot is real, not a glitch).

Then they layered in context. For every day, they computed calendar features: month, weekday, weekend status, season (Winter runs December–February up here), and Norwegian national public holidays — Christmas/New Year, Easter, summer holidays, and the peculiar "bridge days" that Norwegians take to stitch together long weekends. They pulled daily temperature records for Trondheim from the MET Norway Frost API, computing heating-degree days and a temperature-based charging-demand multiplier that scales expected demand upward in the cold.

The behavioral core is a hybrid of two machine-learning techniques. A Conditional Tabular Generative Adversarial Network (CTGAN) — a type of AI where two neural networks compete, one generating fake data and one trying to catch it, until the fakes become statistically indistinguishable from reality — learned the joint distribution of energy delivered, plug-in duration, arrival and departure times, user type (private vs. shared), garage, month, weekday, season, holiday status, and temperature. Alongside it, separate Kernel Density Estimation (KDE) models — a way of smoothing observed data into continuous probability curves — were fitted for each season to preserve the empirical "tails" of behavior, the unusual-but-real long sessions and extreme events that a neat statistical model might wash out.

Fleet growth came from real registration data. Using Statistics Norway (SSB) registered electric-car stock for Trondheim, they computed an annual growth multiplier $G_y={\text{EV}}_y/{\text{EV}}_{2019}$, where 2019 is the baseline year set to 1.00. This multiplier scales the number of daily sessions, not individual session energy — a subtle but important distinction: electrification in this model means more cars charging, not each car charging harder. For years beyond available statistics, they projected conservative, medium, and high adoption trajectories to handle uncertainty through 2030. The medium scenario is the core released dataset.

Finally, and most interestingly, they imposed physics. Real chargers have power limits, and a naive generative model will happily create a session that delivers 30 kWh in two hours — physically impossible on a 7.2 kW AC charger. So they enforced the constraint that delivered energy $E\le P_{\text{charger}}D$, capping energy at 7.2 kW times the plug-in duration. Rather than deleting infeasible sessions (which would distort the session-count structure), they capped the energy and recomputed the average power. This "physical feasibility correction" is what makes the synthetic data trustworthy for downstream power-flow models, which care a great deal about whether a claimed load could actually exist.

The result is validated by comparing the synthetic sessions against the real ones on every measurable axis — correlation heatmaps, scatter envelopes, and marginal probability densities all confirm that the clone behaves like the original, with one intentional difference: the synthetic data extends the variance to model a decade of change.

What They Found

The validation numbers tell the story of fidelity. The observed mean delivered energy per session is 12.73 kWh; the synthetic mean is 12.69 kWh — a hair's-breadth apart. Medians agree even more closely (9.06 vs. 9.43 kWh). Plug-in durations match too: 11.50 hours observed, 11.69 synthetic. These aren't incidental matches; they mean the generator learned that Norwegian EV drivers plug in and leave their cars connected for long stretches, often overnight.

The behavioral signatures are preserved in the correlations. Delivered energy shows a weak positive correlation with plug-in duration ($r=0.11$ observed vs. $r=0.14$ synthetic) — longer plug-in time allows more energy, but the weakness reflects that much of that time is idle parking, not active charging. A strong negative correlation between plug-in duration and average charging power ($r=-0.46$ observed vs. $r=-0.50$ synthetic) captures the physical reality that long connections draw modest average power (think: trickle charge all night) while short "top-up" sessions run at high power.

Arrival times cluster in the late afternoon — peak density between 16:00 and 18:00 — reflecting the classic pattern of Norwegians returning home from work and plugging in. Departures concentrate in the morning. This is the daily rhythm of a residential charging ecosystem, and the synthetic dataset reproduces it faithfully.

Behavioral Fidelity: Observed vs. Synthetic Session Statistics

Label	Value
Observed	12.73
Plug-in duration (h)	11.5
Arrival time (h)	16.82
Departure time (h)	13.05

At the aggregate level, the numbers tell a story about scale and risk. On average, each hour carries 42.33 kWh of charging demand across all active sessions. But averages are not what stress a grid — peaks are. And the peak hourly aggregated demand in the medium-adoption scenario is a striking 1,559.87 kWh. The authors are careful to note this maximum "reflects regional future EV demand rather than a transformer-limited or site-constrained capacity" — in plain terms, this is the raw, unmanaged load that would arrive if everyone charged whenever they wanted. It's the number a distribution-grid planner needs to see before deciding how much transformer capacity to build.

Aggregate Hourly Charging Load: Mean vs. Peak (Medium Scenario, 7.2 kW)

Label	Value
Peak hourly demand (kWh)	1,559.87
Mean hourly demand (kWh)	42.33

One deliberate design choice shapes the whole dataset: it is activity-based, not a continuous time series. Hours with no allocated EV charging energy simply don't appear among the 76,993 records. This is a feature, not a flaw — it means the dataset reflects real charging events rather than padding a timeline with thousands of empty zeros, and it keeps the data compact enough to feed into reinforcement-learning environments and transformer-loading models directly.

The energy conservation between the session-level logs (256,826 sessions) and hourly aggregates is exact — a validation check that sounds dry but is actually the difference between trustworthy data and noise. Total cumulative energy across the 11-year horizon is approximately 3.26 GWh.

The Dataset by the Numbers

Label	Value
Total sessions	256,826
Hourly records	76,993
Cumulative energy (GWh)	3.26

Why This Changes Things

To see why this matters, consider the stakes in the Background section of the paper. Unmanaged EV charging doesn't just use energy — it concentrates it. When thousands of cars plug in simultaneously after a cold evening commute, the aggregated spike can exacerbate peak demand, push voltage deviations, overload lines and transformers, and congest the distribution network. These are not theoretical concerns. The research literature the authors draw on shows that EV charging demand can vary seasonally by up to 16% across regions and increase winter monthly consumption by up to 30% in cold climates (Jack et al., cited in Taraghi Nazloo & Musilek, 2026).

This is precisely why the cold-climate dimension matters so much — and why Trondheim is the perfect subject. Ambient temperature substantially influences both vehicle energy consumption (batteries are less efficient and need heating in the cold) and therefore charging demand. The dataset integrates daily temperature dependencies and cold-weather scaling factors, making it a specialized benchmark for cold-climate EV load forecasting. A dataset calibrated to a warm climate would systematically underestimate winter demand; this one is built for the reality that Norway, and much of the world's northern latitudes, actually faces.

There's also something genuinely forward-looking buried in the growth-modeling. The researchers didn't just extrapolate one trajectory and call it done. They parameterized three — conservative, medium, and high EV-adoption scenarios — so that planners can stress-test their grids against different futures. This is the difference between asking "how much capacity do I need?" and "how much capacity do I need if adoption accelerates?" Given that the actual trajectory of electrification is unknowable a decade out, scenario-based planning is the intellectually honest approach, and the dataset's structure (an adoption-scenario identifier per record) makes switching between futures trivial.

The most subtle contribution, though, is the physical feasibility correction. Synthetic data is only as good as its constraints, and a dataset full of physically impossible charging sessions would mislead every downstream model that consumes it. By enforcing the 7.2 kW AC power limit — and by adjusting rather than deleting infeasible sessions — the authors preserve the behavioral structure while guaranteeing that every record could plausibly exist on real hardware. For transformer-loading analysis and power-flow modeling, where a single implausible load can corrupt an entire simulation, this is the difference between a toy dataset and an engineering tool.

This matters because the dataset's intended users are not just academics. Distribution system operators need realistic hourly load profiles to plan feeder capacity. Energy managers need them to design smart-charging algorithms that shift load away from peak hours. And researchers building reinforcement-learning control systems — where an agent learns when to delay or advance charging to flatten the grid's load curve — need a decade of realistic demand to train against. The paper explicitly positions the dataset as a "cold-climate benchmark for distribution-grid impact assessment, transformer-loading analysis, EV charging-demand forecasting, charger-capacity planning, energy-management optimization, and the development of data-driven smart-charging control strategies."

What's Next

Of course, a decade of projected data has limits, and the authors are admirably candid about them. The generative models were trained on only 14 months of history, and the pipeline cannot anticipate long-term shifts in battery chemistry or capacity, the arrival of vehicle-to-grid (V2G) technology where cars feed power back into the grid, real-time tariff structures that change when people choose to charge, or changes in driver routing behavior. The growth model scales macro-level session counts but doesn't simulate individual decisions, public-versus-private charger infrastructure expansion, or the spatial redistribution of charging across urban zones.

The dataset also represents unmitigated demand — the raw load before any smart-charging control kicks in. The 7.2 kW per-charger hardware limit is enforced, but secondary network constraints (transformer kVA capacity, feeder current limits, phase unbalance) are deliberately left out so downstream studies can apply their own bounds. And it's calibrated to Trondheim specifically: its climate, its EV adoption trajectory, its calendar. Applying it to a region with different weather, lower baseline penetration, or a different grid architecture requires re-calibration against local data. It's a benchmark, not a universal panacea.

None of these caveats diminish the contribution. In fact, they're a feature of good scientific practice — knowing exactly what your tool can and cannot do is what makes it usable. The genuinely exciting part is what this opens up: the ability to test smart-charging strategies not against one lucky season of data, but against a decade of realistic demand that includes cold winters, holiday spikes, and a fleet that keeps growing. When the grid operators of tomorrow need to know whether their transformers can survive a cold Norwegian evening in 2030, they'll have somewhere to look.

Every energy transition is ultimately a question of timing and infrastructure: can the grid absorb what people plug in, and at what hour? This quiet, technical dataset is a piece of that answer — built honestly from real behavior, extended thoughtfully into an uncertain future, and constrained by the physics we can't avoid. That's not just data science. That's the groundwork for a city deciding it can trust electricity to carry it through the cold.

<｜DSML｜tool_calls> <｜DSML｜invoke name="write_paper_digest"> <｜DSML｜parameter name="title" string="true">They Cloned a Decade of EV Charging to Protect a Norwegian City's Grid