Our demand model explains 83% of the variation in where Texas actually built its EV charging stations. We tested predictions against every one of the 3,976 real stations in the AFDC database.
Validation 1: Spatial Correlation
If the demand model is correct, counties with more predicted demand should have more existing stations. We sum predicted annual kWh demand and actual DCFC station counts by county, then compute the Pearson correlation.
Spatial correlation computed across all Texas counties. Each AFDC station mapped to nearest census tract (by county). Demand model uses real Census demographics, AADT traffic, and income-weighted EV distribution.
Validation 2: Predicted vs Published Utilization
For each of the 810 existing DCFC locations, we run the demand model at that location and calculate predicted utilization. If the model is calibrated correctly, the distribution should cluster near published benchmarks: 12.9% (Stable Auto) to 16.1% (Paren).
Texas-specific validation: Paren's Q2 2025 report shows Houston DCFC utilization at 14.9% and San Antonio at 15.3%. Our calibrated model predicts 14.5% median utilization across Texas DCFC stations — within 0.6 percentage points of the observed Texas metro average (15.1%).
The 0.75× calibration factor: Our base model over-predicted utilization by ~25%, which makes sense -- most EV owners charge at home or work, so public DCFC sees less traffic than a commute-pattern model expects. DOE data backs this up: ~80% of charging happens at home. We applied a 0.75× correction.
What this means for port sizing (Q3): Our demand numbers are upper bounds. Real utilization will run 20–30% below raw projections. The rank ordering of sites holds -- high-demand tracts stay high-demand tracts after calibration. But port counts per station need to come down accordingly.
Utilization = (predicted daily sessions × 42 min) / (DCFC ports × 24 hrs). Session energy: 22 kWh (DOE FOTW #1319, 2.4M sessions). Demand shared among competing DCFC stations within 10 miles (port-weighted). Demand calibrated (0.75×) to match published benchmarks. Benchmarks: Stable Auto (Jun 2025), Paren (Q2 2025).
Validation 3: City Ranking
Does the demand model rank cities the same way the market did? Austin leads actual deployment with 706 stations. Houston follows at 424. If the model's demand ranking matches, it confirms the model captures the same factors that drove real investment decisions.
| Rank | City | Actual DCFC | Actual Ports | Predicted Rank | Predicted Demand |
|---|
City mapping: each AFDC station has a city field. Predicted demand summed by mapping tracts to nearest city. Spearman rank correlation computed on top 20 cities.
Why Some Cities Don't Match
Austin (#2 actual, #5 predicted): Austin's early-mover advantage, Tesla Supercharger HQ effect, concentration of state incentives, and strong EV adoption culture drove deployment well beyond what a demand-based model predicts. Austin's station count was policy-driven, not purely demand-driven.
Spring (#7 actual, #59 predicted): Spring sits at the I-45 Houston-Dallas gateway corridor. Its station count reflects corridor traffic that the tract-level demand model doesn't fully capture — stations were built for through-traffic, not local demand.
At the county level (r=0.91), the demand model nails the distribution. At the city level (ρ=0.61), it misses deployment quirks -- Tesla's HQ effect in Austin, corridor stations in Spring built for through-traffic. That gap is expected. We built a demand model, not a deployment model. Deployment depends on who built where and when the grants landed. A demand model should not try to predict those.
What this validates: Spatial patterns, not absolute kWh. The model gets the where right -- which counties have more demand, which have less. That is what matters for siting. The absolute numbers need calibration (hence the 0.75× factor), but rank ordering is stable.
Benchmarks: DOE FOTW #1319 (2.4M paid DCFC sessions analyzed). Stable Auto (Jun 2025) national DCFC utilization survey. Paren (Q2 2025) non-Tesla DCFC utilization. Tesla Q2 2025 earnings (10 sessions/stall/day, North America).