Retail Roots: Planting the Right Stores with Smart AI Soil

TL;DR for operators

Location is the product. For EV charging networks, the technical hardware matters, but the business pain usually begins somewhere less glamorous: the charger is not where drivers actually pause, pass through, or need confidence.

The paper behind this article proposes a data-driven system for recommending EV charging station locations in New South Wales by fusing EV GPS trajectories with existing and approved charger data, LGA boundaries, routes, altitude, fire-risk maps, and points of interest.¹ Its core move is not “AI magically finds the best charger sites”, because apparently we still have to live in the physical world. The useful move is narrower and more operational: use historical trip density to find candidate demand clusters, then constrain those clusters with geospatial features that make a site more plausible.

The authors generate recommendations for 93 fast chargers and 103 destination chargers, visualised against 263 existing fast chargers, 785 existing destination chargers, and approved stations in the NSW master plan. They also provide two case studies: one where recommendations align with existing or approved stations, and another where the system identifies two candidate sites in areas without existing or approved stations.

For business readers, the lesson is not that this is a finished infrastructure investment engine. It is better understood as a decision-support layer. It can help governments, utilities, developers, and charging-network operators ask better first-round questions: where is observed EV movement dense, where are gaps visible, which candidate sites sit near roads or useful POIs, and which LGAs deserve a more detailed planning conversation?

The boundary is equally important. A trajectory-driven model inherits the geography of current EV adoption. If inland NSW has sparse historical EV trips, the system will naturally recommend fewer sites there. That may reflect current usage, but not necessarily future demand. Before money moves, the missing layers still matter: electricity supply, grid connection cost, charger type, load capacity, population growth, land access, utilisation economics, and policy objectives.

A charger is a retail site with a power cable

Think about a new store. The first question is rarely “Which shelf layout is optimal?” It is usually “Should this store exist here at all?”

EV charging has the same site-selection problem, only with more cables, more permitting, and a stronger ability to embarrass planners in public. A charger that is beautifully engineered but badly located is still a stranded asset with branding.

The paper’s practical importance sits in this retail-like framing. Charging stations must be reachable, useful, safe enough, aligned with local geography, and close to where demand actually materialises. The authors do not begin with an abstract optimisation objective or a theoretical grid model. They begin with movement: two months of NSW EV GPS trajectories. That matters because trip traces give a revealed-behaviour signal. Drivers may say they want a charger in one place; their routes often say something more useful.

But raw movement is not enough. A dot on a trajectory map is not a development site. It may be off-road, far from services, exposed to risk, administratively awkward, or simply irrelevant once existing and approved chargers are considered. The paper’s mechanism is designed to move from “drivers pass here” to “this might be a plausible candidate location”.

That distinction is the whole article. The work is not a prophecy machine. It is a geospatial filtering pipeline.

The mechanism: from GPS traces to constrained candidate sites

The system has three broad stages.

First, it cleans and organises EV GPS trip data. Each trip record represents a detailed driving journey. These records become the behavioural base layer: where EVs have actually moved across NSW.

Second, the system fuses this base layer with several planning layers:

Third, it applies a modified DBSCAN clustering process. DBSCAN is useful here because the geography is messy. EV trips are not evenly distributed across neat squares, and LGA boundaries are irregular. A grid may be administratively tidy, but drivers have a rude habit of following roads instead of spreadsheets.

The authors therefore cluster EV trip points within each LGA. The enhancement is that the clustering does not treat every point identically. The search radius and minimum-point threshold can be adjusted according to constraints such as altitude, POI proximity, and route information. In plain business terms: the model does not merely ask, “Where are many EV points close together?” It asks, “Where are many EV points close together in a way that still looks geographically usable?”

That extra step is where the paper’s business relevance lives. A density cluster by itself is only demand evidence. A constrained density cluster becomes a planning candidate.

The system is not choosing sites; it is creating a shortlist worth arguing about

A common misunderstanding would be to treat the output as an investment decision. It is not.

The output is a recommendation map. That is valuable, but it is not the same as a bankable deployment plan. The system can show that a location is plausible from trajectory, access, POI, LGA, altitude, and fire-risk perspectives. It does not show that a site has adequate grid capacity, attractive lease terms, strong future utilisation, or acceptable connection cost.

This distinction is not pedantry. It changes how the system should be used.

The right business use is therefore staged. Use the model to reduce search costs and create a defensible candidate list. Then pass that list into feasibility analysis: grid studies, land availability, permitting, traffic forecasts, charger mix, capex, opex, and utilisation modelling.

In other words, let the AI do the first sift. Do not let it sign the cheque.

What the evidence actually shows

The paper’s experimental section is mainly a visual and case-study demonstration, not a benchmark contest. That is appropriate for a system paper, but it affects how strongly the results should be interpreted.

The authors visualise existing, approved, and recommended charging stations using Folium. Against the NSW master plan data, they report 263 existing fast chargers and 785 existing destination chargers. Their system recommends 93 additional fast-charger locations and 103 destination-charger locations.

That is the main evidence: the pipeline can ingest multi-source geospatial data and produce interpretable, map-based recommendations at NSW scale.

The figures and cases serve different purposes:

There is no reported ablation showing, for example, how recommendations change if POIs are removed, if altitude is ignored, or if LGA boundaries are not used. There is no formal comparison against an alternative optimisation method. There is no utilisation validation against future charging demand.

That does not make the paper weak. It makes it a planning-system demonstration rather than a predictive-performance paper. The useful question is not “Did it beat a leaderboard?” The useful question is “Does the pipeline turn fragmented planning data into a map that operators can inspect and challenge?” On that question, the paper is much more compelling.

The coastal concentration is a result and a warning

One of the more important observations is also the easiest to misread: both existing and recommended stations are more concentrated along coastal LGAs than inland ones.

The authors explain this as a consequence of more developed cities near the East Coast having greater demand. That is plausible and consistent with a trajectory-driven method. If EV trips are denser near developed coastal cities, DBSCAN will find more candidate clusters there.

For operators, this has two meanings.

The first is tactical. If the goal is near-term utilisation, historical EV trip data is exactly the kind of signal one wants. It points toward places where current demand is visible, not merely hoped for.

The second is strategic. If the goal includes regional coverage, future adoption, tourism corridors, emergency resilience, or equity of access, then historical trip density becomes incomplete. Sparse inland recommendations may reflect underdeveloped infrastructure and low current EV adoption, not low future need. A model trained on today’s movement can quietly reinforce today’s infrastructure pattern. That is not bias in the dramatic sense; it is just yesterday wearing a dashboard.

The authors acknowledge this boundary directly: lack of inland historical EV trip data limits recommendations for western inland LGAs, and supplementary data such as population density could help reflect future rural demand.

That is the correct interpretation. The model is strong where current behaviour is informative. It is weaker where the planning objective is to create future behaviour.

Why POIs and roads matter more than they look

The paper’s POI constraint may sound like a small implementation detail. It is not.

Charging is dwell-time infrastructure. Drivers need somewhere to spend the charging interval. For fast charging, that may be food, toilets, fuel-station-style services, tourism stops, or retail. For destination charging, the POI logic is even more direct: charging is tied to the reason for staying.

The system uses POIs such as fast food, fuel stations, and tourism locations as references for aligning recommended charging sites. The authors note that recommended stations in the urban zoom view are located on road segments, benefiting from POI constraints because POIs are usually near roads.

This is where the paper quietly becomes a retail-site-selection paper. A charger does not merely need electrical access. It needs behavioural adjacency. The best charger location is often not the mathematically pure centre of a trip cluster; it is the practical nearby place where a driver can stop without feeling punished.

That is the “smart AI soil” in the title. The algorithm plants candidate stations where demand density meets usable surroundings. It is a modest claim, but a useful one.

The LGA layer turns a model output into a governance object

The LGA boundary layer is not just cartographic decoration. It changes the unit of interpretation.

Infrastructure decisions are rarely made by one omniscient statewide planner with unlimited capital and perfect patience. Local governments, transport agencies, utilities, and private operators all view the problem through different jurisdictions and constraints. By iterating within LGA boundaries, the system can produce recommendations that are easier to discuss locally.

That matters for deployment. A statewide heatmap may be impressive, but a local council needs to know what falls inside its administrative boundary. A charging operator needs to know which approvals and stakeholders are involved. A utility needs to understand where grid planning must be sequenced.

This is one of the paper’s underrated strengths. It does not only generate candidate dots. It embeds them in a governance map.

The business value is cheaper diagnosis, not automatic deployment

The clearest ROI pathway is not “build every recommended charger”. That would be a strangely expensive way to misunderstand a research prototype.

The value is earlier in the funnel:

1. Reduce the search space for candidate sites.
2. Make infrastructure gaps visible against existing and approved stations.
3. Create LGA-specific planning views.
4. Add first-pass accessibility and risk context.
5. Give stakeholders a shared map for discussion before detailed feasibility work begins.

For charging-network operators, this can support market-entry screening. For utilities, it can help identify areas where demand signals may deserve grid-capacity investigation. For government agencies, it can support policy discussions about coverage, regional access, and resilience. For property owners and retail partners, it can identify where charging might complement existing foot traffic.

The cost saving is cognitive and procedural before it is financial. Bad infrastructure projects often begin with vague maps, political preference, and a heroic Excel file. A fused geospatial system does not remove judgement, but it gives judgement something better to argue with.

What still has to happen before this becomes an investment model

The paper is careful enough to name several missing layers in future work. Those layers are not optional for real deployment.

The most important missing layer is electricity supply. A site can be perfect from a traffic and POI perspective and still be painful if the nearby distribution network cannot support the load without expensive upgrades. The authors specifically mention electricity supply near the Sydney area as future data needed to consider charging capacity and maximum acceptable load.

The second missing layer is charger-type classification. The paper reports recommendations for fast and destination chargers, but future work includes using existing charger types and urban functional area divisions to improve charger classification. In business terms, “where should a charger go?” is only half the problem. “What kind of charger should go there?” is the part that decides capex, dwell time, pricing, and utilisation.

The third missing layer is future demand. Historical EV trajectories are useful, but infrastructure has to anticipate adoption curves, housing patterns, tourism, fleet electrification, and policy incentives. Population density is one proposed supplement, especially for rural areas where current trajectory data may be thin.

The fourth missing layer is economics. The paper does not model land cost, connection cost, charger utilisation, maintenance, pricing, competitive response, or payback period. That is fine for the paper’s scope. It is fatal only if a reader tries to pretend the map is a spreadsheet.

A practical operator framework

A sensible organisation would use this kind of system as the first layer in a multi-stage planning process:

The paper delivers the first four layers reasonably well as a prototype. It points toward the fifth. It does not claim to solve the sixth.

That is a clean division of labour. The model should not be punished for not doing everything. It should be judged by whether it improves the part it actually touches.

The strategic lesson: physical AI needs operational humility

There is a recurring mistake in applied AI: assuming that because a model produces a recommendation, the recommendation is a decision. This is especially dangerous in physical infrastructure, where errors become concrete, expensive, and occasionally visible from the highway.

The better lesson from this paper is more disciplined. AI can fuse messy spatial data, reveal demand patterns, and make candidate locations inspectable. It can turn movement traces into a shortlist. It can help planners compare observed demand with current infrastructure. It can surface gaps that deserve human attention.

But the final decision still belongs to a wider operating system: grid engineers, councils, commercial teams, landowners, policy makers, and drivers with their inconvenient human habits.

That is not a weakness. It is the correct architecture. Infrastructure planning should not be automated end-to-end just because a map has nice markers.

The paper’s contribution is a useful middle layer: not raw data, not final deployment, but structured spatial intelligence. For operators, that is often where the real bottleneck sits. Not in knowing that chargers are needed. Everyone knows that. The hard part is knowing which candidate sites deserve the next hour of expensive human attention.

Cognaptus: Automate the Present, Incubate the Future.