EvoXXLTraffic: a sensor-evolving, multi-decade PeMS traffic dataset

EvoXXLTraffic extends the XXLTraffic dataset family with yearly sensor catalogs, persistent station IDs, full-year 5-minute flow matrices and geographic graph snapshots. It covers nine California PeMS districts and up to 27 years (1999–2025), for a total of 213 district-years.

Code, preprocessing notebooks, benchmark protocols and results are on GitHub: https://github.com/cruiseresearchgroup/TSAS26-EvoXXLTraffic (release tsas26-r1.1).

Contents

path	content	size
full_year/<district>/<year>.npz	Full-year flow matrix of every district-year, stored as built by the preprocessing notebooks. See "Full-year matrices" below.	~137 GB
full_year/<district>/graph/<year>_adj.npz	Thresholded geographic-proximity graph of that year's catalog (key x, N × N).	~3.6 GB
full_year/catalog.json	Per district-year: sensor IDs in column order, start time, 5-minute interval, numbers of slots and sensors, and SHA-256 of both files.
caches/<district>/<year>/data.npz, graph.npz	The corrected benchmark tasks used in the TSAS article: the first 31 days of each year, with explicit target and input masks. See "Benchmark caches" below.	~50 GB
coordinates/	Sensor coordinates in column order for every cached district-year, the source snapshot of each, no-look-ahead (dated) coordinates, and tools to build unthresholded distances or any kernel.	small
metadata_snapshots/dXX/	The 3,523 dated PeMS station-metadata files the coordinates come from.	~0.5 GB
tfnsw/	Supplementary hourly data from Transport for NSW, outside the nine districts and the benchmark. See "Supplementary: tfNSW" below.	~0.4 GB
SHA256SUMS, SHA256SUMS.full_year, SHA256SUMS.tfnsw	Checksums of every file.

Years per district:

District	Years	District	Years	District	Years
PEMS03	2001–2025	PEMS06	2005–2025	PEMS10	2006–2025
PEMS04	2001–2025	PEMS07	2001–2025	PEMS11	1999–2025
PEMS05	2005–2025	PEMS08	2001–2025	PEMS12	2002–2025

Processing pipeline

1. Raw records. These are PeMS station 5-minute daily files per district from the Caltrans Performance Measurement System. The raw archive used here is linked from the GitHub README.
2. Yearly catalog. The sorted station IDs present on an anchor day form the catalog. The anchor is the first available day of a district's first year, and 1 January in later years. Daily records are restricted to this catalog. Sensors that appear later in a year enter the next year's catalog. Station IDs persist across years, so a sensor can be followed through time.
3. Full-year flow matrix (full_year/<district>/<year>.npz, key x). Total Flow is aligned to a regular 5-minute grid that runs from the start time to the end of the year. Notebooks: xxltrafficdata/pemsXX_build_eac_data.ipynb on GitHub.
4. Graph (graph/<year>_adj.npz). This is a Gaussian kernel on Haversine distances, exp(−d²/σ²), where σ is the population SD of positive pairwise distances. Weights below 0.1 are removed and the diagonal is zero. Coordinates come from the dated metadata snapshots; a station without coordinates gets the mean location (flagged in coordinates/).
5. Training windows. The original notebooks also cut EAC-format training windows (FastData) from the first 31 days. These windows are a few hundred GB, are fully reproducible from step 3 with the notebooks, and contain backward-filled targets. They are therefore not uploaded.
6. Corrected benchmark caches (caches/). For the article, every yearly task was rebuilt from the daily records with explicit masks and forward-only input filling (benchmark_v2/build_caches.py on GitHub). All reported experiments use these caches.

Full-year matrices

• x has shape slots × sensors (float32).
• Row t is start + 5 min × t, and the year ends at 31 December 23:55.
• Columns follow sensor_ids in full_year/catalog.json.
• In the cached years, the first 31 days equal the benchmark targets wherever those are observed.

Important: the stored matrix is filled. Each sensor's series is forward-filled, then backward-filled, and remaining gaps are set to 0.

• It therefore cannot distinguish a missing record from a true zero.
• Backward filling copies later values into earlier gaps.
• Do not use these values as forecasting targets or as model inputs without accounting for this.
• For leakage-free evaluation, use the masked benchmark caches, or rebuild masks from the raw records.

Benchmark caches (caches/)

Each data.npz holds one yearly task of 8,928 slots (31 days):

key	content
y	Raw targets.
observed	Target accepted: the record is present and the flow is finite.
x	z-scored, forward-filled inputs.
available	Input slot holds a real value.
sensor_ids, timestamps	Column IDs and slot times.
bounds	[0, 5356, 7142, 8928]: train/validation/test.
metadata	Provenance JSON.

Other details:

• The quality fields record_present, flow_reported and percent_observed are kept.
• Hashes equal the data_sha256/graph_sha256 in the GitHub run manifests.
• PEMS04/2001 is part of the archive (full_year/) but not of the benchmark, because its first 31 days have no accepted targets.

Supplementary: tfNSW

tfnsw/ holds yearly hourly traffic matrices built from the permanent counting stations of Transport for NSW (Australia), the non-PeMS source of the XXLTraffic family. It is not part of the nine districts or 213 district-years above, and no corrected benchmark result of the TSAS article uses it.

path	content
tfnsw/<year>.npz	Flow matrix (key x, hours × nodes), 2006–2026. 2026 is partial (January to April).
tfnsw/graph/<year>_adj.npz	Geographic-proximity graph of that year's nodes (key x, N × N).
tfnsw/meta/<year>_nodes.csv	Column order: idx, node, station_id, direction, lat, lon.
tfnsw/catalog.json	Per year: start, 60-minute interval, numbers of slots, nodes and stations, share of zeros, SHA-256 of the matrix and graph.

• Nodes: one node is a station and direction, so a two-way station gives two nodes. Years have 185–847 nodes (99–463 stations).
• Flows: hourly counts of all vehicles, or light plus heavy vehicles where the all-vehicle class is absent.
• The matrices are filled: in each year, every node's series is forward-filled, then backward-filled, and remaining gaps are set to 0. There is no observation mask, so a value here does not mean a count was recorded, and the share of zeros differs by year (2–21%). The publisher also notes data-quality and sensor-availability issues in these datasets.
• Graphs: thresholded Gaussian kernel of Haversine distances between nodes (σ is the standard deviation of the non-zero distances, threshold 0.1, zero diagonal). The two directions of a station are at distance 0.

Caveats

• Graphs describe retrospective geographic proximity, not road connectivity or travel time.
• Some stored graphs use coordinates from snapshots dated after the catalog anchor. The dated_* coordinates avoid this.
• Catalog size counts stored columns, not sensors with valid data at every slot.
• Published flows may include upstream PeMS imputation.

License and attribution

The data products in this repository are released under CC BY-NC 4.0 (LICENSE). The underlying records and station metadata come from Caltrans PeMS (https://pems.dot.ca.gov/), whose terms continue to apply. The records in tfnsw/ come from Transport for NSW, whose Traffic Volume Counts data are published under Creative Commons Attribution (https://opendata.transport.nsw.gov.au/dataset/nsw-roads-traffic-volume-counts-api); they are processed here as described above, and Transport for NSW does not endorse this repository. Please cite the EvoXXLTraffic article (https://arxiv.org/abs/2605.29768) and XXLTraffic (SIGSPATIAL 2025, https://doi.org/10.1145/3748636.3762762).