🌙 Summer Night — overnight heat risk in European cities

A static site showing overnight "feels like" (Wet Bulb Globe Temperature) forecasts for Paris and other major European cities, with an emphasis on nights where the feels-like temperature never drops below 20 °C or 25 °C — the nights that are hardest on human health.

Live site: https://mlgh.net/summer-night/ (also https://mlglobalhealth.github.io/summer-night/ — the org's custom Pages domain)

What it shows

For each city and each night of the coming week (overnight window = 21:00 → 09:00 local time, labeled by the evening's date):

• Feels-like minimum and the hour it occurs, plus the air-temperature minimum
• Hours at or above 20 °C / 25 °C feels-like, shown as the ensemble median with a 10th–90th percentile range (e.g. 4 h (2–6)) — the uncertainty comes from the 40-member ICON weather ensemble
• The last two nights of observed temperature (marked "observed"), before tonight and the forecast week, so you can see the trend the forecast is continuing
• A 12-hour sparkline of the overnight curve, and a 7-day hourly chart (observed portion in grey, forecast in blue, split at "now")
• Row highlighting: orange when feels-like stays ≥ 20 °C all night, red when it stays ≥ 25 °C all night
• A vulnerable-group / mortality banner flagging runs of consecutive nights with no overnight relief, and a "no relief" badge on each such night
• An all-cities overview grid (cities × nights, colored by severity)

"Feels like": apparent temperature

"Feels like" is the Steadman apparent temperature (the Australian Bureau of Meteorology shade formulation):

AT = Ta + 0.33·e − 0.70·ws − 4.0
e  = (RH/100) · 6.105 · exp(17.27·Ta / (237.7 + Ta))       (vapour pressure, hPa)

from air temperature Ta (°C), relative humidity RH (%) and 10 m wind speed ws (m/s). The shade formula has no solar-radiation term, which makes it the correct choice overnight. On humid nights it reads slightly above air temperature; in dry, breezy air slightly below. We compute it identically for the deterministic forecast and for every ensemble member, so the uncertainty ranges are internally consistent.

Why apparent temperature, not WBGT / wet-bulb?

An earlier version used nighttime WBGT. A literature review (see docs/RESEARCH.md) changed this:

• WBGT is built for daytime, in-the-sun physical exertion and is only the best-predicting index for heat mortality in a few tropical countries. It is not validated for overnight sleep or for temperate-European mortality.
• For heat mortality in temperate Europe, plain air temperature predicts about as well as any humidity index (Lo et al. 2023, 39 countries; Armstrong et al. 2019, 445 cities). Where a humidity-inclusive index helps at all in Northern / Eastern Europe, it is apparent temperature — so that is what we use.

The two outcomes, and their thresholds

The site answers two different questions with two different signals:

• Sleep / comfort — a per-night measure. Sleep degrades continuously with warmth (no sharp cut-off), but as benchmarks a feels-like that stays ≥ 20 °C all night is a warm, restless night and ≥ 25 °C all night is oppressive. This drives the per-night table and the hours-above-threshold columns.
• Vulnerable-group mortality — a multi-night measure. Heat deaths in the elderly are driven by consecutive nights with no overnight recovery, not single hot nights: the 2003 Paris study found multi-day minimum (night) temperature predicted elderly deaths while daytime temperature did not, and a 2025 study across 34 European countries found back-to-back day-and-night "compound" heat carried >2× the mortality risk of daytime-only heat. We flag runs of nights with no overnight relief — the feels-like minimum never dropping below 20 °C.

Thresholds are absolute (20 / 25 °C) for readability. The epidemiology actually favours location-specific / percentile thresholds because temperate populations acclimatize (22 °C is routine in Rome, alarming in London); moving the mortality flag to a per-city percentile is the most defensible future refinement.

Data source

Open-Meteo — free for non-commercial use, no API key:

• Forecast API (api.open-meteo.com/v1/forecast): best-estimate hourly temperature and relative humidity, 8 days, in each city's local timezone
• Ensemble API (ensemble-api.open-meteo.com/v1/ensemble, icon_seamless): 40 ensemble members (temperature, humidity, wind) used to compute the uncertainty range on hours-above-threshold and the probability of a no-relief night

Two requests per city per update (22 total), well within Open-Meteo's free tier.

Each city is shareable via a URL hash, e.g. https://mlgh.net/summer-night/#rome or https://mlgh.net/summer-night/epi/#rome.

Three views

• Everyday view — / (https://mlgh.net/summer-night/). "Will you sleep badly and drag through work tomorrow?" Per-night temperature, hours-above-threshold, the last two observed nights, and a comparison of tonight to last night.
• Epi view — /epi/ (https://mlgh.net/summer-night/epi/). The public-health angle: a heat-mortality signal (consecutive no-relief nights), tonight benchmarked against each city's climatology percentiles (acclimatization-aware), and each country's historical summer excess mortality.
• Skill view — /skill/ (https://mlgh.net/summer-night/skill/). Forecast verification: the logged forecast for each night (by lead time) vs what was observed, for hours ≥ 20° and ≥ 25°. Accumulates over time — data/skill.json is appended on every forecast run.

Repository layout

index.html                  everyday site (plain HTML/CSS/JS, no build step, no dependencies)
assets/style.css, app.js    shared styles + everyday-site rendering & SVG charts
epi/index.html              epi site
epi/epi.css, epi/app.js     epi-specific styles + rendering (joins the 3 data files)

data/forecast.json          hourly forecast + per-night summaries (cron, ~3-hourly)
data/climatology.json       per-city ERA5 percentiles + hot-nights-per-year (monthly)
data/mortality.json         per-country summer excess mortality from Eurostat (weekly)

scripts/update_forecast.py  Open-Meteo forecast → apparent temp + night summaries (stdlib)
scripts/build_climatology.py Open-Meteo ERA5 archive → per-city climatology percentiles
scripts/update_mortality.py Eurostat weekly deaths → summer excess mortality
scripts/cron_update.sh      forecast cron: fetch → commit → push (lock + log)
scripts/cron_data.sh        slow-data cron: mortality (weekly) / climatology (monthly)

Reproducing from scratch

git clone git@github.com:MLGlobalHealth/summer-night.git
cd summer-night
python3 scripts/update_forecast.py   # writes data/forecast.json (needs Python ≥ 3.8, stdlib only)
python3 -m http.server 8000          # open http://localhost:8000

Deployment: GitHub Pages

The site is served straight from the main branch root — no build step.

1. Push the repo to GitHub.
2. Repo Settings → Pages → Build and deployment: Source = Deploy from a branch, Branch = main, folder = / (root). Or via CLI:

   gh api -X POST repos/<owner>/<repo>/pages -f 'source[branch]=main' -f 'source[path]=/'

3. The site appears at https://<owner>.github.io/<repo>/ within a minute or two. Every push to main (including the cron job's data commits) redeploys automatically.

Automation: cron on a Linux server

The update script runs on any machine with git push access to the repo (SSH key or token) and Python 3. Every 3 hours is a good cadence — Open-Meteo model runs update roughly that often:

crontab -e
# add:
17 */3 * * * /path/to/summer-night/scripts/cron_update.sh          # forecast, every 3h
30 4  * * 1 /path/to/summer-night/scripts/cron_data.sh             # mortality, weekly (Mon)
40 4  1 * * /path/to/summer-night/scripts/cron_data.sh climatology # + climatology, monthly

The forecast (cron_update.sh) is the frequent one. cron_data.sh refreshes the epi datasets: Eurostat mortality weekly, and the heavier ERA5 climatology monthly (pass climatology to rebuild it — otherwise only mortality is fetched).

The wrapper:

• takes a lock so overlapping runs are skipped,
• git pull --rebase first so pushes don't conflict,
• runs the fetch; on total failure it exits without touching data/forecast.json (the site keeps showing the last good forecast),
• on partial failure it publishes the cities that succeeded and keeps each failed city's previous data (marked "older data" in the UI),
• commits and pushes only when the data actually changed,
• logs to logs/update.log (kept to the last ~2000 lines).

Check on it with tail logs/update.log.

Caveats

• Outdoor vs indoor: we forecast outdoor temperature, but sleep and health depend on the bedroom. Homes without air-conditioning often stay warmer than the outside air overnight and cool more slowly, so indoor conditions can be worse than these numbers suggest — and nearly all of the threshold evidence in the literature is measured indoors. Treat outdoor feels-like as a proxy that may understate risk.
• Forecasts, especially beyond 3–4 days, are uncertain — that's what the ranges are for. This is not an official heat warning; follow national meteorological services.
• Apparent temperature here is computed at a single grid point per city; urban heat islands can make real neighbourhoods warmer.
• Thresholds are absolute for readability; the epidemiology favours location-specific (percentile) thresholds because of acclimatization.
• The ensemble (ICON seamless) is coarser than the deterministic forecast, so the central "hours" estimate and the range can disagree slightly.
• All times are local to each city; the overnight window is 21:00–09:00.

See docs/RESEARCH.md for the literature review behind these choices, with sources.