Lesson 4 · 13:30–15:00

When should we warn?

Satellites can't see tomorrow, so the alert uses air temperature (2 m above the ground): will tomorrow's forecast high cross Bengaluru's heat gauge?
First, each day's highest (tmax) and lowest (tmin) air temperature since 1991 from ERA5, Europe's hour-by-hour rebuild of past weather on ~25 km squares (it reads about 1.5 °C below the city's weather station on summer afternoons). See it, get it. Climate Pulse maps ERA5 for the whole world, day by day, and Open-Meteo gives any place's daily values as a CSV, which is what we download for Bengaluru (35 years in 0.2 MB):
era5 = datapack.table("era5_daily")  # each day's highest and lowest air temperature (ERA5, via Open-Meteo), °C
era5[["tmax", "tmin"]].tail()
tmax tmin
time
2026-09-26 29.8 20.7
2026-09-27 30.7 19.4
2026-09-28 30 21.8
2026-09-29 28.8 21.9
2026-09-30 29.1 21.5
How this file was made · making it reads 35 years of daily ERA5 from Open-Meteo, a few seconds

The chart reads a small prepared file (0.2 MB): your copy in data/ if you have one, otherwise it downloads once. To make it yourself from the original data, run uv run wardheat datapack build --step climate --force. This is the code that runs:

def era5_daily(start: str = "1991-01-01", end: str = "2025-12-31", point=CITY_POINT, model: str = "era5") -> pd.DataFrame:
    """Daily ERA5 history for one point (°C, %)."""
    r = requests.get(
        "https://archive-api.open-meteo.com/v1/archive",
        params=dict(latitude=point[0], longitude=point[1], start_date=start, end_date=end,
                    daily=",".join(DAILY_VARS), timezone="Asia/Kolkata", models=model),
        timeout=120,
    )
    r.raise_for_status()
    return _to_frame(r.json())


def step_climate(force=False):
    from . import climate

    if datapack.exists("era5_daily") and not force:
        return
    d = climate.era5_daily("1991-01-01", "2026-09-30")
    d.to_parquet(_out("era5_daily"))
    _log(f"climate: {len(d)} days of ERA5")

40 °C never comes

In 35 years no day in Bengaluru reached 40 °C, so India's 40 °C rule would never warn the city. So we use one Bengaluru gauge for every day of the year: a forecast high from 33 °C is warm, 34 °C hot, 35 °C very hot and 36 °C extreme; 33 °C is hotter than 9 in 10 days of 1991–2020, and each level is one degree more:
normal = warning.climatology(era5, base=(1991, 2020))  # the average high for every date, 1991-2020

# a level for every day since 1991, from its highest air temperature
levels = warning.levels(era5.loc["1991":], lines=(33, 34, 35, 36))
Here is 23 April 2026, this year's hottest day, on the gauge:
Every March–May since 1991 on the same gauge:
2024 had the longest run of very hot or extreme days (16 in a row). Hot-or-worse days each March–May: 12.9 a year on average in 1991–2000, 18.2 in 2011–2020.

Tomorrow's forecast

Open-Meteo's free forecast gives the coming days' highest air temperature, read straight off the gauge:
forecast = climate.forecast(days=7)  # today and the next 6 days, as Open-Meteo gives them
week = warning.levels(forecast, lines=(33, 34, 35, 36), normal=normal)
week[["tmax", "tmax_normal", "in words"]].round(1).rename(columns={
    "tmax": "high °C", "tmax_normal": "average high for the date °C", "in words": "level"})
high °C average high for the date °C level
time
2026-10-14 31.7 27 below the alert line
2026-10-15 29.5 27 below the alert line
2026-10-16 26.1 27 below the alert line
2026-10-17 26.9 26.9 below the alert line
2026-10-18 27.3 26.9 below the alert line
2026-10-19 25.7 26.9 below the alert line
2026-10-20 26.4 26.9 below the alert line

369 ward alerts

Each ward's alert combines the city's level with the ward's risk class from lesson 3:
Tomorrow's high is… below 33 °C 33 °C+ warm 34 °C+ hot 35 °C+ very hot 36 °C+ extreme
High-risk ward – take care heat alert emergency emergency
Medium-risk ward – take care take care heat alert emergency
Low-risk ward – – take care heat alert emergency
wards_risk = indices.ward_risk(indices.ward_table())["risk"]  # High / Medium / Low, as in lesson 3

tomorrow = week.iloc[1]  # row 0 is today
this_year = era5.loc[str(era5.index[-1].year), "tmax"]
hot_day = levels.loc[this_year.idxmax()]  # the example above: this year's hottest day (ERA5)
alerts = pd.DataFrame({
    "tomorrow": warning.ward_alerts(int(tomorrow.level), wards_risk),
    "hot_day": warning.ward_alerts(int(hot_day.level), wards_risk),
})
alerts.apply(lambda alert: alert.value_counts()).fillna(0).astype(int)  # how many wards get each alert
tomorrow hot_day
Emergency 0 111
Heat alert 0 258
No alert 369 0

Sending the alerts

Every warned ward gets a message for its officers and residents; when tomorrow has no alerts, the cell shows what this year's hottest day would have sent:
day, when = ("tomorrow", tomorrow) if (alerts.tomorrow != "No alert").any() else ("hot_day", hot_day)
advice = {
    "Take care": "drink water often and stay out of the sun from 12 noon to 3 pm.",
    "Heat alert": "stay out of the sun from 12 noon to 3 pm, drink water and ORS, check on older neighbours and outdoor workers.",
    "Emergency": "stay indoors from 12 noon to 3 pm, drink water and ORS, call 108 for anyone who collapses in the heat.",
}
send = alerts.loc[alerts[day] != "No alert", [day]].rename(columns={day: "alert"})
send.insert(0, "ward", load_wards().set_index("ward_uid")["label"])
send["message"] = send.alert.map(lambda a: f"{a} for {when.name:%d %B}: {when.tmax:.0f} °C air temperature; " + advice[a])
send.sort_values(["alert", "ward"]).head(10)
ward alert message
ward_uid
26 Central 26 · Vannarpet Emergency Emergency for 23 April: 35 °C air temperature; stay indoors from 12 noon to 3 pm, drink water and ORS, call 108 for anyone who collapses in the heat.
27 Central 27 · Ambedkarnagar Emergency Emergency for 23 April: 35 °C air temperature; stay indoors from 12 noon to 3 pm, drink water and ORS, call 108 for anyone who collapses in the heat.
28 Central 28 · Neelasandra Emergency Emergency for 23 April: 35 °C air temperature; stay indoors from 12 noon to 3 pm, drink water and ORS, call 108 for anyone who collapses in the heat.
32 Central 32 · SilverJubilee Park Ward Emergency Emergency for 23 April: 35 °C air temperature; stay indoors from 12 noon to 3 pm, drink water and ORS, call 108 for anyone who collapses in the heat.
34 Central 34 · D.V Gundappa Ward Emergency Emergency for 23 April: 35 °C air temperature; stay indoors from 12 noon to 3 pm, drink water and ORS, call 108 for anyone who collapses in the heat.
39 Central 39 · Venkat Reddy Nagara Emergency Emergency for 23 April: 35 °C air temperature; stay indoors from 12 noon to 3 pm, drink water and ORS, call 108 for anyone who collapses in the heat.
45 Central 45 · K.R Market Emergency Emergency for 23 April: 35 °C air temperature; stay indoors from 12 noon to 3 pm, drink water and ORS, call 108 for anyone who collapses in the heat.
46 Central 46 · Cheluvadi Palya Emergency Emergency for 23 April: 35 °C air temperature; stay indoors from 12 noon to 3 pm, drink water and ORS, call 108 for anyone who collapses in the heat.
48 Central 48 · Azad Nagar Emergency Emergency for 23 April: 35 °C air temperature; stay indoors from 12 noon to 3 pm, drink water and ORS, call 108 for anyone who collapses in the heat.
49 Central 49 · Kasturbha Nagar Emergency Emergency for 23 April: 35 °C air temperature; stay indoors from 12 noon to 3 pm, drink water and ORS, call 108 for anyone who collapses in the heat.
369 messages in all: 258 Heat alert, 111 Emergency.

The same idea, live

The US weather service runs this every day: its HeatRisk map colours the whole country green to magenta for the next 7 days by comparing the forecast with local records, and heat.gov tells people what to do.

Try it

  1. Click 2016 and 2019 in the calendar.
  2. Move the gauge one degree up:
higher = warning.levels(era5.loc["1991":], lines=(34, 35, 36, 37))  # change the lines here
_hot = lambda lv: (lv["level"] >= 2).groupby(lv.index.year).sum()
pd.DataFrame({"33/34/35/36": _hot(levels), "34/35/36/37": _hot(higher)}).loc[2010:].rename_axis("hot or worse days")
33/34/35/36 34/35/36/37
hot or worse days
2010 20 4
2011 0 0
2012 18 1
2013 23 2
2014 35 18
2015 11 0
2016 34 23
2017 29 1
2018 3 0
2019 22 2
2020 7 1
2021 2 1
2022 3 1
2023 8 0
2024 37 26
2025 1 0
2026 17 2
Next steps: set the gauge from hospital records, add humidity, warm nights and hot spells, and street sensors. Thank you!
Made with open data and open-source Python · marimo · Altair