EV charging infrastructure operator · 2024
AI demand forecasting for EV charging infrastructure
A predictive platform that forecasted EV charging demand by combining public datasets, private telemetry, and computer-vision detection of existing infrastructure.
Impact
- Site selection shifted from reactive prioritization driven by anecdote and partner referrals to a data-driven model with a defensible market map underneath it.
- The detection layer surfaced private and semi-private charging locations that public registries don't cover, materially changing how target geographies were sized.
Business challenge
Capital planning for charging-network expansion was constrained by a market map that was both incomplete and stale. Public datasets cover the well-known public sites; they miss private fleets, semi-private depots, and recently-deployed infrastructure that changes the competitive picture for any given corridor. Demand forecasts were built on top of that incomplete map, so site-selection decisions were absorbing data error and forecasting error compounded.
Approach
The platform was built in two halves, intentionally separable.
The forecasting half combined gradient-boosted models for tabular drivers (vehicle registrations, traffic counts, demographics, utility-level constraints) with neural networks for spatial-temporal patterns the boosted models couldn't capture cleanly. Public and private data were unified into a feature store that treated provenance as a first-class attribute, so any forecast could be traced back to the inputs that produced it.
The detection half used CNN-based image classification on Google imagery to find charging infrastructure the public registries missed. Detections were reconciled with private operator data through a confidence-weighted merge, producing a single canonical map of the network as it actually existed, not as it appeared on file.
Keeping the two halves decoupled mattered. Forecasting could run on whatever map was current; detection could improve the map without forecasting needing to retrain. That separation made each pipeline iterate at its own cadence and made root-cause attribution clean when a number looked wrong.
Impact
Deployment planning moved from a reactive, referral-driven workflow to data-driven prioritization with an auditable basis for each ranking. The detection layer changed the operator's market map in several target geographies: locations that didn't show up in any public dataset were now visible, and the competitive picture shifted accordingly.
Hard numbers around capital efficiency and downstream commercial impact aren't public.