A charger in the wrong place does two kinds of damage. It sits idle where few EVs pass, and the busy area that needed it gets queues and a strained local transformer. Most discussion of public charging in Delhi NCR starts with sites and hardware. I think it has to start one step earlier, with a forecast of where the vehicles will be.
That is the subject of my part-time PhD at Amity University Haryana. The committee approved my synopsis on 17 November 2026, so this is work in progress. Here is what the data already shows and how I plan to get from registrations to charger sites.
What the VAHAN data already shows
VAHAN is the national vehicle registration database. In October 2026 I pulled monthly registrations for January 2021 to August 2026 from a public extract of the VAHAN dashboard, and the 2025 vehicle-category split from the dashboard itself. All numbers in this section come from that 2026 pull.
In that period 6,85,398 EVs were registered in Delhi NCR, 8.7% of 78,46,750 vehicles. The average hides large gaps between cities:
- Delhi: 11.2% of registrations were electric
- Noida: 9.0%
- Ghaziabad: 8.6%
- Gurugram: 3.8%
- Faridabad: 2.8%
Delhi and Faridabad share a border, yet the EV share in Delhi is about four times higher. A single plan for the whole region would miss that.
The vehicle segment matters even more than the city. In 2025, commercial three-wheelers (e-rickshaws and e-autos) were 81 to 88% electric in Delhi, Noida and Ghaziabad, but only 35% in Gurugram. Private two-wheelers and cars were between 2% and 7.5% electric in every city. Gurugram's low overall share is mostly a three-wheeler story. Three-wheelers and cars also charge at different places and times, so I forecast each segment separately.
The supply side is hard to see
Counting existing chargers is harder than counting vehicles. The Delhi government's list has 891 sites, 573 for charging and 318 for battery swapping. OpenStreetMap shows only 160 stations across the whole of NCR. Open data is dense for Delhi and thin for Gurugram, Noida, Faridabad and Ghaziabad, so building one combined list of existing stations is part of the work.
There are also gaps that no download will fill. India has no open charging-session data, and transformer data are not public. Where I cannot measure something, I will set it through low, medium and high scenarios and say so.
Why a good curve fit is not enough
My first attempt was the obvious one. I fitted Bass diffusion curves to cumulative registrations, and they matched history very closely, with R² above 0.99. The trouble was the market size. When the model estimates it freely, it comes out only 1.2 times the EVs already registered since January 2021 in Noida and Gurugram. Taken at face value, that would mean EV growth there is nearly finished at a share below 10%, which is not believable.
So a close fit to the past says little about the ceiling. In the main study the market size will be fixed through low, medium and high scenarios based on total vehicle registrations, and I will compare the Bass model with a Gompertz curve.
Three steps: adoption, demand, siting
The framework links three questions that are usually answered separately.
- Adoption. Fit Bass and Gompertz curves for each RTO and vehicle segment, then use regression and XGBoost to explain why cities differ, using factors such as population density, chargers and the share of commercial vehicles.
- Demand. Divide NCR into zones of about 1 km² on the H3 hexagonal grid, share each RTO forecast among zones by population and activity, and estimate daily charging energy in kWh for each zone.
- Siting. Score candidate land with GIS and AHP, choose sites with a p-median model solved by a genetic algorithm, size each site with an Erlang C queue so waiting time stays below a target, and check that the chargers fit within the local transformer limit.
To see whether this beats a simpler approach, I will compare the plan with one that spreads chargers on a uniform grid, for 2030 and 2035, on demand covered, waiting time, number of chargers and transformer violations. Sites that come up in every scenario get recommended first.
I have tested the forecasting part on open data. On the UrbanEV dataset from Shenzhen, which gives hourly charging energy for 275 zones, gradient boosting cut the average error by about 30% compared with a forecast that simply repeats the same hour yesterday. The pilot ran on an ordinary laptop in under a minute. Those energy values are estimates made by the dataset authors, so I used them only to compare methods, not to measure real demand.
What I would tell a planner today
If you are placing chargers in NCR this year, plan for each city and vehicle segment separately. Start with commercial three-wheelers in Delhi, Noida and Ghaziabad, and do not treat Gurugram's low overall EV share as low demand everywhere.