Most infrastructure playbooks start with a simple assumption: nearby assets compete. Gas stations have successfully operated this way for decades. Proximity equals competition. Stations close to each other share demand patterns, compete for the same drivers, and respond to each other’s pricing. With AI and machine learning (ML), that often turns into models that watch nearest competitors, respond to their pricing, and try to capture market share within a geographic radius. 

That assumption has carried over into EV charging. When a new EV charging station opens near your site¹, the instinct is to treat it as competition. The closer it is, the more it threatens your volume. 

But EV charging isn't fueling. A gas station visit takes 4 to 8 minutes. An EV charging session last anywhere between 20 to 45 minutes. Drivers make fundamentally different decisions about when and where to charge. They aren't choosing between two nearby chargers in the moment; the choice is often made earlier, based on where they plan to spend time. A site near a grocery store, a workplace, or a travel corridor serves different intent, even if they’re a few blocks apart. 

This suggests that proximity may be a weak signal for modeling competition. To assess this, I shifted my focus from geography to demand behavior, that is, comparing stations based on how they behave rather than how close they are. 

The original goal of this work was narrow: forecast charging demand at sites with minimal operational history i.e., a few-shot forecasting problem. In the process, however, the analysis revealed a more fundamental structure in charging behavior. 

I worked with a proprietary dataset of around 8,000 public DC fast charging sites across the top five charging networks by utilization. The data consisted of session-level records, including energy sold and start and end times, covering anywhere from a few months to multiple years per site. The task was straightforward: given a site with limited history, predict demand over the next seven days. 

A global time-series model, specifically a Temporal Fusion Transformer, performed reasonably well. It picked up shared temporal patterns and handled seasonality by averaging over very different behaviors. 

But the demand spread across sites was obvious even before training. A charger on a travel corridor over Thanksgiving behaves nothing like a downtown workplace charger. That raised a more useful question: can we surface these behavior patterns systematically? And more importantly, can forecasting itself be the tool that uncovers them?

To make this concrete, I represented each stations demand as a weekly utilization profile i.e., average hourly demand from Monday through Sunday. Each site was also summarized with a set of statistical features that capture how demand distributes across the calendar, including periodicity and outliers. Clustering those representations with k-means resulted in groups of similar behavior. The harder question is how many groups actually matter. 

For each candidate k, I trained k “expert” models, one per cluster, and evaluated forecast performance. Performance peaked at 12; below that, distinct behaviors were merged, and above it, the models overfit and lost transferability. That was the signal: twelve distinct site archetypes that capture dominant demand patterns while enabling robust few-shot inference.

Once the clusters stabilized, the patterns were easy to read. The sites were attached with context such as amenities and land use features to help interpret them, but the structure was already visible in the demand curves as seen in the figure below. To read the figure, follow a single curve across the week: peaks indicate periods of high activity, whereas troughs reflect low usage. The shape of the curve captures when demand occurs, independent of scale. Comparing curves reveals distinct behavioral patterns, such as weekend spikes (A2) and steadily increasing weekday usage (A8). 

These patterns appear across charging networks². What stands out is that similarity ignores geography. Two sites thousands of miles apart can follow nearly identical demand patterns if they serve the same underlying behavior, while neighboring sites may differ significantly. 

The key insight is that these groupings emerge from usage patterns alone, not proximity or surrounding context. 

These archetypes also help reframe how competition is defined. Since stations close to each other may belong to different archetypes, they can serve distinct demand pools, peak at different periods, and respond differently to pricing changes, and therefore may not be competing at all.

The results indicate that proximity is a convenient but weak proxy for competition and demand, while demand behavior serves as a more informative signal. 

This finding has implications across the stack. 

Pricing: Pricing decisions often still follow geography: anchor to the nearest site and adjust. But as we've seen, not all nearby stations compete and price elasticity varies by archetype. A commuter corridor responds differently to price changes than a leisure site or an airport adjacent charger. The same price increase can reduce utilization in one case and have little effect in another. Anchoring to nearby stations can push prices in the wrong direction, either leaving revenue on the table or degrading the user experience. Using archetypes makes pricing conditional on behavior instead of distance, and provides a better starting point. 

Siting: Archetypes also change how new locations are evaluated. The question shifts from "what do nearby stations look like" to "which demand pattern will this site fall into." That shifts attention to land use, dwell time, and traffic patterns. A site next to retail, even in a different city, is more comparable to other retail sites than to the closest charger down the street. This helps avoid overestimating demand in locations that look attractive on a map but don’t match a proven usage pattern. 

Planning: At the grid level, these patterns matter for load shape. Different archetypes stress infrastructure in different ways, such as steady high utilization versus sharp peaks around commuting hours or holidays. Planning with archetypes gives a more realistic view of aggregate demand. It helps utilities and operators anticipate when and where load will concentrate, rather than relying on averages that smooth over important differences. 

More generally, this work highlights the assumptions in complex systems, and how often they're inherited from contexts where they may no longer apply.

This shows up repeatedly in applied AI. When performance saturates, the default response is to scale: more parameters, more compute, and more data. But often, the bigger unlock comes from representation: finding the right way to describe the system before trying to model it. In this case, once sites were grouped by behavior, better predictions followed naturally.

This work did not begin with the intent to identify structure, but rather to solve a forecasting problem. The structures emerged because they were a real feature of the underlying system. This provides a useful perspective for understanding AI systems more broadly: when a structure consistently improves performance, it likely reflects a real signal. And it can uncover things that can help and change the way people make decisions and shape policies.

A detailed report is published here.

Roam is the virtual workspace our team relies on to stay connected across time zones. It makes collaboration feel natural with shared spaces, private rooms, and built-in AI tools.

Roam’s focus on human-centered collaboration is why they’re our Premier Partner, supporting our mission to connect the builders and leaders shaping the future of AI.

The AI Collective is a community of volunteers, made for volunteers. All proceeds directly fund future initiatives that benefit this community.

Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.