matrosovdani/moscow-parking-occupancy

Moscow Parking Occupancy Dataset

4.6 million half-hourly occupancy snapshots for 210 municipal parking lots in Moscow, continuous since 31 March 2025 and refreshed every Monday.

Long, dense, public occupancy series are rare: most published work on parking occupancy prediction still leans on the UCI Parking Birmingham set, which covers October–December 2016 only.

Configs

Config	Rows	What it is
occupancy	~4.6 M	Snapshots every 30 minutes
parking_spots	210	Static metadata: coordinates, capacity, nearest metro

from datasets import load_dataset

occ = load_dataset("matrosovcmtn/moscow-parking-occupancy", "occupancy", split="train")
spots = load_dataset("matrosovcmtn/moscow-parking-occupancy", "parking_spots", split="train")

Schema

occupancy — primary key (time, parking_id)

Column	Type	Description
time	timestamp[ns, UTC]	Snapshot timestamp
parking_id	int32	Foreign key to parking_spots.id
free_spaces	int32	Free spaces (common + accessible)
free_handicapped_spaces	int32	Free accessible spaces
occupancy_rate	float64	Percent 0–100 of common spaces occupied

occupancy_rate is computed as (common_total - common_free) / common_total * 100 over declared common capacity only — it is not 1 - free_spaces / total_spaces.

parking_spots — id, external_id, name_ru, name_en, address_street_ru, address_street_en, subway_ru, subway_en, latitude, longitude, total_spaces, common_spaces, handicapped_spaces, last_free_at, feed_silent.

Read this before modelling: 39 lots are not what they look like

19% of the Moscow lots report a constant zero free spaces. That looks like "completely full". It is not — the city registered these lots but does not publish their occupancy, or the sensor died.

Measured on 2026-09-20: 39 of 210 lots reported no free space at all in the previous 30 days; 19 never reported one in the entire history. One 257-space lot last showed a free space in May 2026, a 143-space lot at VDNKh in July 2025. A 606-space lot does not stay full for five months.

Kept in the data on purpose, but flagged:

spots = spots.to_pandas()
live = set(spots.loc[~spots.feed_silent, "id"])
occ = occ.filter(lambda r: r["parking_id"] in live)   # drops ~4.8% of rows

last_free_at lets you rebuild the flag for any as-of date instead of trusting ours.

Separately, the data is censored at the boundaries: 25.2% of all rows are exactly 100 and 7.8% exactly 0 — mostly genuine, lots really do fill up. Treat the target as bounded.

Caveats

• Municipal lots only. Malls, private lots and unpaid curbside are not in the Moscow public API and are not here.
• Reported, not observed. Values come from the upstream API; occasional zero-spikes and stale values are preserved rather than cleaned.
• Gaps exist during upstream outages.
• No PII. Aggregate counts only — no vehicle, plate or driver data.

Licence and citation

Data: CC BY 4.0. Original values come from Moscow Department of Transport public parking data; this is a derivative work that aggregates and republishes them. Keep attribution to both.

@misc{matrosov2026moscowparking,
  author       = {Matrosov, Danil},
  title        = {Moscow Parking Occupancy Dataset},
  year         = {2026},
  howpublished = {\url{https://github.com/matrosovcmtn/moscow-parking-occupancy}}
}

Canonical source and full documentation: github.com/matrosovcmtn/moscow-parking-occupancy · Maintained by ParkOut