T2M · T2M_MAX · T2M_MIN
Air temperature
°C at 2 m
Drives all three indices
Feeds the Rothfusz Heat Index, the WBGT wet-bulb and globe terms and the UTCI approximation. It also anchors the 3-day and 7-day anomaly features of the forecast.
THERMORA / Planetary heat intelligence
The planet is heating.Our responsestarts earlier.
Transforming environmental signals into actionable heatwave early warnings and human thermal stress insights, helping communities understand heat risk and prepare for extreme conditions.
Extreme heat · Thermal stress · Early warning
Scroll to begin
Area of interest 28.6692° N, 77.4538° E
Ghaziabad, Uttar Pradesh
02Environmental signals
Every heat crisisbegins as a signal.
THERMORA reads daily point observations from NASA's POWER service for every ward centroid. Four of its parameters shape human heat stress directly. Each one has a precise, documented role in the calculation.
RH2M
Relative humidity
% at 2 m
Limits evaporative cooling
Enters the Heat Index regression and scales the wet-bulb estimate, T × (0.45 + 0.55 × RH). Moist air makes the same temperature physiologically harsher.
WS2M
Wind speed
m/s at 2 m
The protective signal
Cools the globe term by 0.5 °C per m/s (capped at 6 m/s) and lowers UTCI by 2.265 °C per m/s above 0.5 m/s. Still air is weighted as added stress.
ALLSKY_SFC_SW_DWN
Solar radiation
W/m², daily mean
Radiant heat load
Converted from MJ/m²/day. Adds up to 8 °C to the black-globe temperature and 12 °C to mean radiant temperature per 1000 W/m², raising stress beyond air temperature.
Surface pressure (PS) and precipitation (PRECTOTCORR) are stored with every observation but do not enter the HTSI formula.
Source: NASA POWER, daily point API, community AG
03Understanding heat risk
A heatwave is a pattern,not a single hot day.
The forecasting stack labels a day as part of a heatwave when temperature stays at or above the location's 90th percentile for at least three consecutive days. It learns that pattern from rolling 3-day and 7-day means, temperature anomalies, humidity trends and the seasonal cycle.
Designed ensemble
Three models, one probability
Tree models read 13 engineered features. The LSTM reads 7-day sequences of temperature, humidity, wind and solar radiation. Their probabilities are blended, with confidence decaying by 4% per forecast day after the first, never below 60%.
XGBoost
45% weight
Random Forest
30% weight
LSTM
25% weight
P(heatwave)
- XGBoost400 trees, depth 5, learning rate 0.05
- Random Forest300 trees, depth 10, balanced classes
- LSTM2 layers, 32 hidden units, 7-day sequences
Running in this deployment
The transparent fallback
The trained model files come from a separate Python training job. Where they are absent, the original code switches to a physically grounded logistic curve on the recent temperature anomaly, and that is what produces every forecast you see here.
P = 1 / (1 + e−(Δ − 3) / 2)
Δ = mean of last 3 days − mean of the 30-day window, P clipped to 1% to 99%
In fallback mode every forecast day receives the same probability, and the three model fields report that same value. No accuracy figure is published for either path.
04Human thermal stress
Temperature tells a story.Human thermal stress tells us why it matters.
The Human Thermal Stress Index (HTSI) folds three established biometeorological indices into one 0 to 100 score, then adjusts it for the person exposed: age, chronic illness, outdoor work and hours in the sun. Heat Index, WBGT and UTCI stay distinct inputs, each with its own scale.
Heat Index ring, 40%WBGT ring, 35%UTCI ring, 25%
Live result
—/ 100 HTSI
No data
- Heat Index
- —°C
- WBGT
- —°C
- UTCI
- —°C
How the score is built
Heat Index
40%
NOAA/NWS Rothfusz regression
Scaled 0 to 100 over 27 to 54 °C
WBGT
35%
Outdoor approximation with solar load and wind cooling
Scaled 0 to 100 over 23 to 35 °C
UTCI
25%
Reduced first-order polynomial approximation
Scaled 0 to 100 over 26 to 46 °C
HTSI = min(100, (0.40 HI + 0.35 WBGT + 0.25 UTCI) × vulnerability), where vulnerability adds 0.15 for age 65+ or 5 and under, 0.15 for chronic illness, 0.10 for outdoor work, 0.03 for women and 0.03 per hour of sun beyond two (up to 0.25).
Risk classes
- Safe0 to 30
- Caution30 to 55
- Danger55 to 80
- Extreme Danger80 to 100
HTSI is a planning signal, not a medical diagnosis or a guarantee of safety.
05Early warning
From a readingto a warning,in six stages.
This is the pipeline as implemented in the platform's code, including where it currently runs on demonstration data and where an external key switches on a live feed.
- 01
Environmental signals
Daily temperature, humidity, wind and solar radiation for each ward centroid. A forward 5-day leg can be added from OpenWeather when an API key is configured.
- NASA POWER daily point API
- OpenWeather 3-hourly forecast, aggregated per day
- 02
Data acquisition
Ingestion requests the last 30 days for a ward, allowing for NASA POWER's two-day latency, converts solar energy to W/m² and stores every day as an observation.
- POST /api/weather/ward/{id}/ingest
- Seeded demo history: 45 synthetic days per ward
- 03
Data processing
Rolling 3-day and 7-day means, the temperature anomaly against the expanding mean, a humidity trend and a cyclic day-of-year encoding form the model inputs.
- 13 tabular features
- 7-day sequences for the LSTM
- 04
Heat-risk assessment
Each ward receives an HTSI score with a driver breakdown, a 5-day heatwave probability with severity and duration, and a hospital surge estimate.
- HTSI 0 to 100 in four classes
- Forecast severity: Low, Moderate, High, Extreme
- 05
Early warning
When a ward's HTSI reaches the alert threshold of 55, the alert engine issues a warning with its recommended action. A refresh job recomputes every ward from its latest observation, designed to run every 30 minutes.
- ALERT_RISK_THRESHOLD = 55
- Alerts expire after 12 hours
- 06
Actionable information
Alerts fan out to people registered in the ward by push, email, SMS and WhatsApp. Each channel sends only when its provider is configured and otherwise logs a dry run.
- Cooling-center finder and health guidance
- District overviews for officers
06Intelligence dashboard
Read straight fromthe platform, right now.
The panel below is not a mockup. It queries the same service as the dashboard and labels where every number comes from, including when that source is demonstration data.
thermora / live readout
connectedHighest HTSI, all wards
Selected ward
—
No data
- Heat Index
- — °C
- WBGT
- — °C
- UTCI
- — °C
Air temperature, last 0 observations
Source: …Forecast engine: …
07The human impact
Heat is felt by people,not by averages.
These are the settings THERMORA was designed to support. They describe intended applications, not verified deployments or measured outcomes.
Urban heat-risk awareness
Dense, paved wards hold heat longer. The map ranks every ward so attention goes where stress is highest.
Outdoor worker safety planning
Exposure hours and outdoor work raise the personal multiplier, and Danger-level guidance calls for work-rest cycles.
Community heat preparedness
Ward-level HTSI and plain-language guidance help residents see when ordinary days become dangerous, and where the nearest cooling center is.
Health guidance · Cooling-center finder
Emergency-response planning
Surge estimates for heat-stroke cases, admissions and ICU demand, alongside staged playbooks, give planners a common picture before a heat event peaks.
Hospital surge · Emergency playbooks
Environmental monitoring
Daily observations, thresholds and alert history per ward make heat stress something that can be tracked over time instead of reported after the fact.
Alerts · Thresholds · Weather history
Climate-data analysis and research
Every formula, weight and threshold is documented, so results can be inspected, challenged and recalibrated against local data.
Methodology · Open API
08About THERMORA
Earlier warnings,clearer reasons.
THERMORA is an extreme-heat early-warning and human thermal stress intelligence platform, built for a Smart India Hackathon problem statement on urban heatwave risk management. It turns meteorological data into ward-level heat-risk intelligence for disaster-management authorities and citizens, starting from a default area of interest in Ghaziabad, Uttar Pradesh.
It is a decision-support tool. It does not issue official meteorological warnings, and its outputs are not medical advice.
Show the source
Every reading is labelled with where it came from and when, including when it is demonstration data.
Explain the score
Each HTSI value ships with its drivers, and every weight and threshold is published on the methodology page.
Never fill gaps
Missing or unavailable data is shown as missing. Nothing on the platform is estimated to make a panel look complete.
Current coverage
States and districts loaded in the platform today. Ward boundaries are approximated as squares around each centroid until official boundaries are attached.
