fumd-ai/krakow-handover-dataset

PID: http://hdl.handle.net/21.15131/2VgDpMDP

FUMD-AI Kraków Vehicular Handover Dataset

AI-ready, per-second traces of vehicles moving through a simulated 5G NR deployment over central Kraków, with every serving-cell change classified and every row labelled with an upcoming-handover flag and the cell the vehicle will settle on.

Each row joins a vehicle's mobility state (position, speed, heading, lane) to its radio state at the same second (serving cell, SINR, CQI, RLC delay/throughput, distance to the serving gNB). The dataset is the input to the FUMD-AI handover forecaster, which predicts the serving cell 1–7 seconds ahead, but it is a general-purpose resource for vehicular-mobility and proactive-handover research.

• Four runs, differing only in traffic density: 900_1, 1000_1, 1200_1, 1400_1 (the number is the simulation's nominal vehicle count; _1 is the seed/repetition).
• 1,408,271 labelled rows across 3,925 vehicles, sampled at 1 Hz over 1,800 simulated seconds per run.
• 27,628 detected serving-cell change events, each classified into one of five cases.
• 9 base stations, ids 1–9, with coordinates included.

At a glance

Run	Rows	Vehicles	Distinct lanes	Warning-window rows (migration 1 or 2)	Change events
900_1	259,918	775	3,486	5,734 (2.21 %)	5,552
1000_1	307,879	887	3,534	6,535 (2.12 %)	6,309
1200_1	373,688	1,045	3,668	7,718 (2.07 %)	7,292
1400_1	466,786	1,218	3,649	8,652 (1.85 %)	8,475
Total	1,408,271	3,925	—	28,639 (2.03 %)	27,628

All four runs cover t = 1 … 1799 s, the same 9 cells, and the same geographic extent (longitude 19.9233–19.9513, latitude 50.0527–50.0707).

Loading

from datasets import load_dataset

# per-second labelled traces (default config)
ds = load_dataset("FUMD-AI/krakow-handover-dataset", split="run_900_1")

# all four densities at once
ds = load_dataset("FUMD-AI/krakow-handover-dataset")   # DatasetDict of 4 splits

# the serving-cell change event log
ev = load_dataset("FUMD-AI/krakow-handover-dataset", "events", split="run_900_1")

# the 9 base-station sites
bs = load_dataset("FUMD-AI/krakow-handover-dataset", "base_stations", split="sites")

Or straight from the files, which is what the FUMD-AI workflows themselves do:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download("FUMD-AI/krakow-handover-dataset",
                       "data/900_1/dataset_labeled_w3.csv",
                       repo_type="dataset")
df = pd.read_csv(path)

Note on metadata/krakow_bs_coords.csv. The file carries a long # comment header describing its provenance. pandas.read_csv(..., comment="#") reads it directly; the base_stations config above is served by the Hub's own parser, which also skips those lines.

Repository layout

data/900_1/dataset_labeled_w3.csv    per-second labelled trace, 37 columns
data/900_1/events_all_w3.csv         serving-cell change event log, 10 columns
data/1000_1/ … data/1400_1/          the same two files per run
metadata/krakow_bs_coords.csv        the 9 base-station sites (id, name, lon, lat)
metadata/krakow_osm_extract.geopackage.zip
                                     the OpenStreetMap extract the SUMO network
                                     was built from — resolves `lane` ids to roads
parquet/                             optional Parquet mirror of data/ (see below)
dataset_info.json                    machine-readable schema + per-run statistics
ro-crate-metadata.json               RO-Crate 1.1 description of this dataset
CITATION.cff                         citation metadata

The CSVs are the authoritative, byte-exact files the published model was trained on; nothing has been re-encoded. parquet/ is a convenience mirror produced by scripts/make_parquet.py — the Hub also auto-converts the CSVs to Parquet for the dataset viewer, so you never need the mirror to browse or query the data.

The labeled config — per-second traces

One row per vehicle per simulated second, 37 columns.

Mobility (from SUMO)

Column	Type	Meaning
t	float	simulation time, seconds; 1 … 1799
veh_id	int	vehicle id, unique within a run
x	float	WGS84 longitude, degrees (~19.9233–19.9513) — not metres
y	float	WGS84 latitude, degrees (~50.0527–50.0707) — not metres
angle	float	heading, degrees clockwise from north, 0–360
speed	float	m/s; 0–19.2 observed
pos	float	metres travelled along the current lane; 0–345 observed
lane	string	SUMO lane id — see the note below
slope	float	road gradient. Constant 0 in every run; the SUMO network is flat
signals	float	turn-signal bitmask; 6 distinct values observed (0, 1, 2, 8, 9, 10)

lane has two shapes. Ordinary road lanes are an OSM way id plus a lane index (-94092847#1_0). Lanes inside junctions are SUMO internal edges, which start with a colon (:4069144706_0_0) and have no OSM counterpart. Internal lanes are 2,396 of the 4,109 distinct ids and about 20 % of all rows — so roughly one row in five cannot be matched to the OSM extract by lane id alone. Use x/y for those.

Radio (from OMNeT++ / Simu5G)

Column	Type	Meaning
Time	float	OMNeT++ timestamp. Duplicates t after the merge; kept for traceability
Object	int	OMNeT++/VEINS module index for the vehicle. Provenance only — not a feature
averageCqiDl	float	mean downlink CQI index, 1–15; fractional because it is an average
distance	float	metres to the serving gNB; 4.7–514 observed
measuredSinrDl	float	downlink SINR measured by the UE, dB; −20.9 … 86.3 observed
measuredSinrUl	float	uplink SINR, dB; −14.8 … 97.7 observed
rcvdSinrDl	float	downlink SINR of received transmissions, dB; −14.3 … 91.9 observed
rlcDelayDl	float	downlink RLC delay, seconds; median 4.3 ms, max 1.95 s
rlcPduDelayDl	float	downlink RLC PDU delay, seconds; coarse (31 distinct values), max 58 ms
rlcThroughputDl	float	downlink RLC throughput in Simu5G's own units, not rescaled; 0–3,144

The three rlc*Dl columns are exactly 0 for vehicles that never carried downlink application traffic — about a third of the rows in 900_1. That is a deliberate fill by the preprocessing workflow, not missing data. Treat 0 there as "no downlink traffic", and do not impute it.

Serving cell, past and future

Column	Type	Meaning
servingCell	int	current serving cell, 1–9
servingCell-1 … servingCell-7	int	serving cell 1–7 seconds in the past. -1 = no history that far back; 0 = not connected
servingCell1 … servingCell7	int	serving cell 1–7 seconds in the future — these are the forecasting targets. Same sentinels: -1 = the vehicle's trace ends before that second

0 is a reserved "not connected" code across all of these columns. It never appears in servingCell itself — 0 rows out of 1,408,271 — and appears in only 191 cells of the look-ahead columns across the whole dataset, which is why the reference model's 10-wide output head effectively predicts 9 classes.

⚠️ The -1 sentinel is not a class. Rows near the start of a vehicle's trace carry -1 in the lags, and rows near the end carry -1 in the leads. A model trained on -1 targets as if they were a tenth cell will learn to predict "trace ended". Drop those rows, or mask those horizon steps.

Horizon	Rows with lead = -1	Rows with lag = -1
±1 s	1,406 (0.10 %)	3,813 (0.27 %)
±4 s	9,286 (0.66 %)	15,442 (1.10 %)
±7 s	17,487 (1.24 %)	27,121 (1.93 %)

⚠️ Leakage. servingCell1 … servingCell7 are look-ahead columns. Using them as model inputs trivially solves the forecasting task. The reference model uses only servingCell and the seven servingCell-k lags.

Labels

Column	Type	Meaning
migration	int	0 steady state · 1 first row of the pre-handover warning window · 2 remaining rows of that window
destination	int	the cell the vehicle will actually settle on — inside a warning window the upcoming stable cell, outside it the current/next stable cell

destination uses destination_mode="freeze_stable", so a short-lived ping-pong blip is never reported as a destination. migration == 1 fires exactly once per actionable handover: 9,574 rows across the four runs, matching the 9,574 actionable events (C1 + C2a + C3) in the event log.

Label distribution:

migration	Rows	Share
0 — steady state	1,379,632	97.97 %
1 — warning-window start	9,574	0.68 %
2 — warning-window continuation	19,065	1.35 %

The class imbalance is the central difficulty of this dataset. A model scored on all rows is scored mostly on steady state; the published forecaster reaches 0.9886 top-1 accuracy on steady-state rows and 0.6283 on warning-window rows at +1 s. Report both.

Labelling parameters (the w3 in the filenames)

Parameter	Value	Meaning
window_s	3.0	how many seconds ahead of a confirmed handover the warning window opens
tol_s	0.1	tolerance: a run counts as stable if its duration ≥ window_s − tol_s
thr	2.9	the resulting stability threshold in seconds
announce_policy	strict	previous and destination run must be stable
handle_bursts	True	handovers preceded by short unstable runs are also detected
destination_mode	freeze_stable	ping-pong blips are not reported as destinations

The original study swept window_s ∈ {2, 3, 4, 5}; only the 3-second labelling is published here. Re-labelling with another window is a single call to fumd_workflow.labeling.annotate_migrations in the preprocessing workflow.

The events config — serving-cell change log

One row per detected serving-cell change, including changes that are not actionable handovers. This is the diagnostic log behind the row labels.

Column	Type	Meaning
veh_id	float	vehicle id (float here, int in the labeled config)
t_change	float	second at which the serving cell changed
from_cell, to_cell	int	cells either side of the change
case	string	classification, see below
policy_src	string	which detection branch fired: strict, strict_burst, strict_burst_scan, strict_startburst
thr, window_s, tol_s	float	the labelling parameters above, repeated per row
note	string	free-text qualifier. Empty for exactly the three actionable cases (C1, C2a, C3) and always populated for C2b and C4, so note.isna() is an equivalent actionable-event filter

case	Meaning	Actionable?	Total
C1_handover_normal	stable → stable, direct transition	✅	5,480
C2a_ABC	stable → short runs → a different stable cell	✅	3,970
C2b_pingpong	stable → short runs → back to the same stable cell	❌ not a real handover	2,474
C3_handover_sin_historico	first stable run reached after only short runs since the start of the recording	✅	124
C4_no_estable	entering a short/unstable run right after a stable one	❌ diagnostics only	15,580

C4 dominates the log (56 %) and C2b adds another 9 %: most serving-cell changes in this scenario are instability, not handovers. Only the 9,574 actionable events produce warning windows in the labeled config. Filtering the event log by case is the intended way to study ping-pong behaviour.

The base_stations config

metadata/krakow_bs_coords.csv gives the 9 sites as cell_id, name, lon, lat. The cell_id values match servingCell / destination exactly.

The layout is a synthetic 3 × 3 uniform grid, ~667 m spacing, over central Kraków; each site is additionally labelled with the nearest landmark (Rynek Główny, Wawel, Kraków Główny station, …) purely for human readability. It is not a real operator deployment and is not derived from RSRP measurements — no such measurements exist for this project.

Scenario and provenance

Area	central Kraków, Poland — bounding box lon 19.918–19.956, lat 50.053–50.071
Road network	OpenStreetMap extract, included as metadata/krakow_osm_extract.geopackage.zip
Mobility	SUMO, fcd-output, 1 Hz
Network	OMNeT++ 6.3.0 + Simu5G 1.4.4 + INET 4.5.4 (5G NR), VoIP downlink application traffic
Base stations	9, synthetic 3 × 3 uniform grid, ~667 m spacing
Duration	1,800 simulated seconds per run (t = 1 … 1799)
Densities	nominal 900 / 1000 / 1200 / 1400 vehicles, one seed each (_1)

Raw SUMO fcd-output XML and OMNeT++ .vec vectors were turned into these tables by the FUMD-AI preprocessing workflow (v1.1.5): parse → lagged position columns → OMNeT++ vector matrix → SUMO/OMNeT id alignment and merge → past-position sentinel fix → migration labelling. The events_all_w3.csv files are that workflow's Step 6 diagnostic output.

Intended uses

• Proactive handover forecasting — predict servingCell1…7 from the mobility and radio history. This is what the reference model does.
• Handover-event detection and classification — predict migration, or reproduce the case taxonomy with a different policy.
• Ping-pong analysis — the C2b / C4 events isolate instability that a naive serving-cell-change count would report as handovers.
• Mobility modelling and spatial analysis — x/y/lane against the OSM extract, with density as a controlled variable.
• Cross-density generalisation — train on one density, test on another. The four runs share a map, a deployment and a labelling policy, so density is the only intended difference.

Limitations and cautions

1. The base-station layout is a synthetic uniform grid. Real deployments are irregular and terrain-shaped. A uniform grid almost certainly makes handovers more predictable than they are in the field, so accuracy measured here is an optimistic bound.
2. No real radio measurements. Every radio column is Simu5G's model output, inheriting that model's propagation, scheduling and mobility assumptions.
3. x and y are degrees, not metres. Any scaler fitted on them is valid only inside this ~2 km × 2 km extent and does not transfer to another city.
4. lane is map-specific, and ~20 % of it is not OSM at all. Ordinary lane ids are OSM way ids from this extract; the rest are SUMO internal junction lanes. Either way an encoder fitted on them is meaningless elsewhere — the Alicante sibling dataset shares zero lane ids with these runs. Per-run vocabularies are 3,486–3,668 ids and their union across the four runs is 4,109, which is exactly the vocabulary size of the reference model's fitted lane encoder.
5. Density is confounded with seed. One run per density (_1), so a difference between 900_1 and 1400_1 mixes the density effect with run-to-run variance. Do not read small differences as density effects.
6. Short vehicle traces. Some vehicles contribute as few as 1–2 rows (median ≈ 275–308). A 6-step window with a 7-step horizon needs 13 rows, so short traces yield no training windows at all — filter before windowing.
7. Severe label imbalance (2.03 % warning-window rows), discussed above.
8. slope is constant 0 and Time duplicates t; Object is a simulator module index. None of the three carries signal — the reference model drops all three.
9. veh_id dtype differs between the two configs (int in labeled, float in events). Cast before joining.
10. Vehicle ids are not comparable across runs. veh_id = 42 in 900_1 and in 1400_1 are unrelated vehicles.
11. Not for real-network deployment as-is. This is a research artefact from a single synthetic scenario.

Relationship to the published model

FUMD-AI/handover-forecaster was trained on all four runs combined: 1,085,658 training windows and 271,759 validation windows, split by vehicle. Its bundled scaler was fitted on 6,513,948 rows (training windows flattened) and its lane encoder on the union of the four runs' lane vocabularies. Reproducing that split from this dataset requires the training workflow's Step 1b (COMBINE_DATASET_PATHS).

Ethics and privacy

Entirely synthetic. No human subjects, no personal data, no real network traces. veh_id identifies a simulated vehicle. The road geometry comes from OpenStreetMap, © OpenStreetMap contributors, available under the Open Database License; the bundled extract in metadata/ is redistributed under those terms, which are separate from this dataset's CC-BY-4.0 licence.

Licence

CC BY 4.0 — use, share and adapt freely, with attribution. The OpenStreetMap-derived extract in metadata/ remains under ODbL (see above).

Citation

@misc{fumdai_krakow_handover_dataset,
  title        = {FUMD-AI Krak{\'o}w Vehicular Handover Dataset},
  author       = {Bernad, Cristina and Filiposka, Sonja and Gilly, Katja},
  year         = {2026},
  howpublished = {Hugging Face Hub},
  url          = {https://huggingface.co/datasets/FUMD-AI/krakow-handover-dataset},
  note         = {SUMO + OMNeT++/Simu5G simulation over central Krak{\'o}w;
                  produced by the FUMD-AI preprocessing workflow}
}

See CITATION.cff for the machine-readable form.

Authors and funding

• Cristina Bernad — 0000-0001-9537-415X
• Sonja Filiposka — 0000-0003-0034-2855
• Katja Gilly — 0000-0002-8985-0639

Funded by the FUMD-AI project, an EOSC GRAVITY Inter Project, Grant Number 25-EOSC-GRV-INTER-013.

Acknowledgements

We gratefully acknowledge Polish high-performance computing infrastructure PLGrid (HPC Centers: ACK Cyfronet AGH) for providing computer facilities and support within computational grant no. PLGINT/2026/019844.

The research work was supported by the Open Science Cloud research laboratory (OSC-LAB) at the Faculty of Computer Science and Engineering (FINKI), Ss. Cyril and Methodius University in Skopje, North Macedonia.

Related resources

• Sibling dataset: FUMD-AI/alicante-handover-dataset — identical schema over central Alicante, with an irregular real-landmark base-station layout
• Model: FUMD-AI/handover-forecaster
• Preprocessing workflow: github.com/FUMD-AI/fumd-ai-preprocessing-workflow
• Training & postprocessing workflow: github.com/FUMD-AI/fumd-ai-training-postprocessing-workflow