Analyse history¶
After a few weeks of collecting, turn the lake into something you can write about. Three questions a reviewer will ask: how complete is the data, what does a typical day look like, and can the stations be grouped by behaviour. Coverage comes first — there is no point clustering rhythms from a station that was offline half the month.
Requirements
Extras: [parquet], [cluster]. Input: a Parquet lake built by the previous step.
What the script does:
- Builds the availability panel with
build_availability_panel. - Reports per-station uptime and the longest gap with
coverage_report(no imputation). - Derives net flow (
calculate_net_flow) and clusters daily rhythms withcluster_diurnal_profiles/label_diurnal_typology.
Run it:
Full source:
"""After a few weeks of collecting: turn the lake into something you can write about.
python 03_analyze_history.py /data/velib_lake --system-id velib --tz Europe/Paris
Three things a reviewer will ask for: how complete is the data, what does a typical day
look like, and can the stations be grouped by behaviour. Coverage first — there's no point
clustering rhythms from a station that was offline half the month.
"""
import argparse
from pathlib import Path
import gbfs_toolkit as gb
def main():
ap = argparse.ArgumentParser()
ap.add_argument("lake", type=Path)
ap.add_argument("--system-id", default="velib")
ap.add_argument("--tz", default="Europe/Paris", help="local tz for diurnal analysis")
args = ap.parse_args()
# convert to local time *before* any daily aggregation, or rush hour lands at the wrong hour
panel = gb.build_availability_panel(args.lake, system_id=args.system_id, target_tz=args.tz)
cov = gb.coverage_report(panel)
print(
f"median uptime {cov['uptime_pct'].median():.0f}% "
f"({(cov['uptime_pct'] > 90).mean():.0%} of stations above 90%)\n"
)
# only trust stations we actually observed most of the time
well_observed = cov.index[cov["uptime_pct"] > 80].get_level_values("station_id")
panel = panel[panel.index.get_level_values("station_id").isin(well_observed)]
typ = gb.cluster_diurnal_profiles(panel, n_clusters="auto", normalize="zscore")
labels = gb.label_diurnal_typology(typ.set_index(["system_id", "station_id"]))
print("station typologies:")
print(labels.value_counts().to_string(), end="\n\n")
flow = gb.calculate_net_flow(panel)
busiest = (
flow.assign(activity=flow["net_flow"].abs())
.groupby("station_id")["activity"]
.sum()
.sort_values(ascending=False)
.head(5)
)
print("most active stations (lower-bound turnover):")
print(busiest.to_string())
if __name__ == "__main__":
main()