Model Validation Report

Can We Predict Where EV Chargers Will Be Used?

Validating a data-driven site selection model against real-world charging station usage in the Paris suburbs.

Saint-Denis & Aubervilliers
Seine-Saint-Denis (93) • Île-de-France • March 2026
Based on Research
Défi n°2 des infrastructures de recharge : le taux d’utilisation
Music & Girard — Mob-Energy, 2023
Read the Paper
0
Charging points
evaluated in the zone
0
With real usage data
for validation
Confirmed
Higher-rated locations
are genuinely busier

What the Theory Predicts

The Mob-Energy research paper (Music & Girard, 2023) argues that many charging stations fail not because of low EV demand, but because of operational traps.

  • Visibility matters most. Stations hidden behind buildings or down side streets get ignored, even in high-demand areas.
  • Parking duration must match. The best locations are where people naturally stop for 30–90 minutes — shopping centers, business districts, errands.
  • Passing traffic generates usage. Stations on routes people drive through, not just live near, outperform isolated locations.

Many charging stations fail not because nobody drives an electric vehicle nearby, but because the station is invisible, the parking behavior doesn’t match the charging speed, or drivers simply prefer to charge at home.

Music & Girard, Mob-Energy Operational Analysis, 2023

What the Data Showed

We measured how strongly each factor predicts real-world station usage. Higher values mean the factor is a better predictor.

Tourism & Transit FlowVisitor & commuter traffic
+0.413
AC Chargers (≤22 kW)Everyday slow chargers
+0.327
Site Visibility & Foot TrafficNear shops, visible from road
+0.217
DC Fast Chargers (50kW+)Highway-speed chargers
−0.150
Model predicts well Model does not predict

How We Tested It

Real-Time Data Collection

We collected live availability data from the French government’s open data API (QualiCharge/IRVE) for every charging station in the Saint-Denis zone over several weeks, tracking when each charger was occupied vs. free.

Model vs. Reality

We compared our model’s location ratings against actual usage patterns. The core question: do stations we rate highly actually get used more? We ranked all 213 stations by both our predicted score and their real occupancy.

Key Finding

Stations rated in the top 25% by our model have measurably higher real-world usage than those rated in the bottom 25%. The model’s rankings match reality — better-rated locations are genuinely busier.

How Each Station Was Classified

Every station falls into one of four categories based on how our model rated it versus how much it’s actually used.

Top Quartile

High score, high usage.
Model confirmed — good site.

Trapped

High score, low usage.
Model over-predicted.

Hidden Gem

Low score, high usage.
Model missed this one.

Bottom Quartile

Low score, low usage.
Model confirmed — poor site.

← Low Model Score  •  High Model Score →   |   ↓ Low Usage  •  High Usage ↑

What We’re Still Working On

Fast Chargers (DC, 50kW+)

Usage of highway-speed chargers is driven by navigation apps and corridor effects, not local demographics. We need a different modelling approach for these.

Other Regions

Validated in Saint-Denis, promising in Valenciennes. Dense inner-Paris suburbs (Montreuil, Créteil) show different patterns we haven’t yet captured.

Precision

The model identifies better and worse locations, but it ranks — it doesn’t forecast exact utilization rates. It’s a compass, not a GPS.

See It Yourself in the App

Open ev.iointegrated.com and expand Polling areas in the left panel.
Click Saint-Denis Aubervilliers — the map zooms to the zone and filters the data.
Click the chart icon next to the zone name to open the evaluation report.
View colored markers on the map: green = confirmed good, red = confirmed poor, amber = over-predicted, blue = hidden gem.
Toggle additional layers to overlay population density, EV adoption rates, and occupancy hotspots.
We have a data-driven method to evaluate charging station locations, tested against real usage data from over 200 stations. It doesn’t replace on-the-ground assessment, but it narrows the search — instead of evaluating 50 potential locations, we can identify which 15 are worth the site visit.