Synthetic data for FTS-NNet reproducibility -- Traffic speed, intensity, hwv intensity, weather, geometry and accidentality
Description
Synthetic dataset ap7_synthetic_5min.csv.gz is a synthetic stand-in for the AP-7 corridor data used in the paper, with the same 39 columns, types, separator (;) and row layout as the real 5-minute input, so every pipeline stage runs on it unchanged. The real records belong to the Catalan Traffic Authority and are not redistributed. No value comes from the real data. Road geometry is drawn from the summary statistics in Table A.2 of the paper; traffic, weather forecasts and crashes come from simple parametric processes (src/data/make_synthetic_data.py). Models trained on it learn the generator, not the AP-7, so their scores say nothing about the paper's results. Committed file vs. real data: kilometre posts 150–159 (10) vs. 120–219 (100); 2 directions in both; periods Jun 2024 and May–Jun 2025 vs. Apr 2024 – Sep 2025; 524,160 vs. 26,323,200 segment-intervals; 0.068 % vs. 0.0567 % crash-stamped. Stages 1 and 2 train on June 2024 and May 2025; June 2025 is held out. Larger files come from the generator, e.g. full size (about 5 GB): python -m src.data.make_synthetic_data --pk-min 120 --pk-max 219 --windows 2024-04-04:2025-10-01 Without --out, it writes to the filename every pipeline stage reads under $AP7_DATA_DIR; scripts/run_demo.sh unpacks the committed file there. Schema (one row per directional kilometre-post segment and 5-minute interval) - dat: interval start (local time); via: highway; pk: kilometre post (km, lower bound of the 1-km segment); sen: direction (dec = south–north, cre = north–south) - anyo, mes, dia, diaSem (0 = Monday), hor, 5min: date and time - 1d_fcst_*, 3d_fcst_*: 1- or 3-day-ahead forecasts, constant within each hour, of temperature, precipitation, snowfall, cloud cover, wind speed and gusts, plus a rain flag (≥ 0.1 mm). The paper uses only the 3-day forecast. - mean_speed (km/h), intTot and intP (total and heavy-vehicle, veh / 5 min), car (lanes), each with an imputation flag - ACCIDENT: 1 inside a logged crash affectation (dilated ±15 min by the pipeline); C_NIVELL_AFECTACIO (level 0–5), F_TEMPS_AFECTACIO (h), F_LONG_AFECTACIO (km) - mob_esp: special-mobility calendar flag; ang_curv, ang_pend_pos, ang_pend_neg: curvature and slopes (deg); segment: road-section ID How the synthetic processes work - Traffic: weekday peaks at 08:00 and 18:00, weekend profile, summer uplift and holidays. Speed falls near lane capacity; congestion spreads to neighbouring posts and persists. Heavy-vehicle share rises at night. - Weather: a hidden hourly series; the forecasts are noisy copies, the 3-day one noisier and missing more rain. - Crashes: rare onsets whose hazard rises with congestion, rain and curvature. Each stamps a queue of 1–4 posts upstream for 15–90 min, setting ACCIDENT and the *_AFECTACIO fields and slowing traffic, so the label describes the crash-affected state (paper, Section 4.5.2).