LoneWolfgang/att-tower-rca

AT&T Tower Dataset for RCA-Triage Simulation

A national map of AT&T cell towers where every tower is labeled with an estimate of how many customers it serves. The customer-weighting is the point of the dataset: it lets a fault simulation tie root cause analysis to customer impact — when a tower or cluster goes down, you can estimate how many people are affected and rank hotspots by population, not just by tower count.

Disclaimer / honesty note. estimated_population_served is a rough proxy, not real subscriber data. It is built from residential population (not commuter load), a flat national market-share assumption, and a nearest-tower apportionment model. AT&T carrier attribution comes from broadcast network codes, not from AT&T. This dataset is intended for simulation, visualization, and training, not for operational or commercial decisions about real subscribers.

Dataset Summary

• Tower locations come from OpenCellID: a cell counts as AT&T by its broadcast MCC/MNC (radio identity), not by who owns the physical mast.
• Customer weighting is produced by apportioning US Census ACS block-group population to the nearest tower(s) via a KD-tree and scaling by AT&T market share.
• Region labels (state, Census quadrant) are assigned by point-in-polygon against real Census boundaries, so towers near state borders are attributed correctly.
• Nearest-city labels support naming hotspot clusters in a downstream dashboard.

Each row is one unique AT&T cell, identified by MCC-MNC-AREA-CELL.

Supported Tasks

• Population-weighted fault triage / RCA simulation — the primary use case: tie a modeled fault to an estimated number of impacted customers.
• Geospatial analysis of AT&T cell-site distribution by state and Census region.

Data Fields

column	type	description
cell_global_id	string	MCC-MNC-AREA-CELL composite cell identifier (primary key)
latitude, longitude	float	Decimal degrees. Estimated unless position_exact.
state	string	2-letter USPS code, point-in-polygon assigned (may be null offshore)
state_name	string	Full state name
quadrant	string	Census 4-region split: Northeast / Midwest / South / West
estimated_population_served	float	Estimated customer-weight (see Methodology)
nearest_city	string	Nearest populated place from a ~30k-city reference list
nearest_city_state	string	State of the nearest city (in-state preferred within an 8 km margin)
nearest_city_km	float	Distance to that nearest city, kilometers
radio	string	GSM / UMTS / LTE. 5G NR rides LTE and is not a separate class.
mcc	int	Mobile country code (310–316 for US)
mnc	int	Mobile network code (carrier identity)
area	int	TAC/LAC
cell	int	Cell ID
samples	int	OpenCellID measurement count (higher = more reliable position)
range	int	Estimated cell range, meters
position_exact	bool	True if GPS-precise (changeable=0), else averaged
is_firstnet	bool	True for FirstNet (313-100) cells riding AT&T's RAN
carrier_note	string	Audit note from the AT&T code table

Methodology

Carrier attribution. Cells are matched to AT&T by an explicit, editable (MCC, MNC) code table covering AT&T's principal commercial codes (310-410, 310-280), legacy codes, and FirstNet (313-100, flagged is_firstnet). MNC↔carrier mappings drift over time, so a few codes carry a "verify" note in carrier_note.

Deduplication. A cell can appear across multiple MCC files or twice within one file. Duplicates on MCC-MNC-AREA-CELL are collapsed, keeping the observation with the most measurements (best averaged position).

Region labeling. Towers are assigned a state by point-in-polygon against the Census 1:500k cartographic state boundaries (not bounding boxes), then mapped to a Census 4-region quadrant.

Population apportionment. For each Census block group, its population is split among its k nearest towers (KD-tree, in an equal-area CRS so "nearest" is in meters). A tower's estimated_population_served is the sum over every block group that picked it, times AT&T market share. Nearest-tower apportionment self-adjusts to density (dense areas → small per-tower territories; sparse → large) and assigns each block group's residents exactly once (no double-counting). Defaults: k=1, market_share=0.45.

City labeling. Each tower gets its nearest populated place from a merged ~30k-city reference, preferring an in-state city when an out-of-state one is only marginally closer (fixes border slips like a Memphis-edge tower snapping to West Memphis, AR).

Source Data

source	role	license
OpenCellID	tower locations, radio type, network codes	CC BY-SA 4.0
US Census ACS 5-year (table B01003)	block-group population	Public domain
US Census TIGER / cartographic boundaries	state + block-group geometry	Public domain
US city reference lists (GitHub)	nearest-city labels	per-source

Considerations for Using the Data

Biases and limitations.

• Population served counts residents, missing daytime/commuter load — downtown cores are under-weighted relative to true peak demand.
• A flat national market share is applied everywhere; real AT&T share varies by region.
• OpenCellID coverage tracks where contributors travel, so the footprint is observation-driven, not a complete network inventory; rural estimates are softer.
• radio reflects documented OpenCellID classes only; this is "where AT&T cells are," not a verified 5G map.

Intended use. Simulation, visualization, and training. Not intended to represent real subscriber counts, real coverage, or operational network state.

Attribution and License

This dataset is derived from OpenCellID, licensed CC BY-SA 4.0. Per the share-alike terms, this derived dataset is also released under CC BY-SA 4.0, and any product using it must visibly credit "OpenCelliD" with a link to https://opencellid.org/ and share derivatives under the same license. Census data is in the public domain.

Citation

@misc{att_rca_triage_dataset,
  title  = {AT&T Tower Dataset for RCA-Triage Simulation},
  note   = {Derived from OpenCellID (CC BY-SA 4.0) and US Census ACS/TIGER},
  url    = {https://opencellid.org/}
}