Methodology & data transparency
What CLIMATIQ knows, how it knows it, and where it is uncertain
CLIMATIQ is a hackathon-stage decision-support prototype. It is not an official meteorological service and never replaces warnings from the India Meteorological Department (IMD) or disaster-management authorities.
What is implemented, approximated or planned
| Capability | Status | Notes |
|---|---|---|
| Historical climate (5 years) | Implemented — real data | ERA5 reanalysis via Open-Meteo for all states; seasonal windows for pilot districts. |
| Live forecast guidance | Implemented — real data | Open-Meteo forecast API (numerical weather prediction), refreshed by ingestion runs. |
| Heat forecasting model | Implemented — statistical baseline | baseline-v1 (no machine learning yet). Adapter interface ready for ML models. |
| Severity classification | Approximation of IMD criteria | Uses a 5-year reference climate instead of IMD's 1991–2020 station normals. |
| Official IMD warnings | Not connected | IMD's API requires registration/approval; the adapter is in place but not configured. |
| Weather stations (IoT) | Ingestion API implemented; stations simulated | No physical hardware is deployed. Demo stations are clearly labelled SIMULATED. |
| AI advisories | Implemented with fallback | LLM provider when configured; otherwise a deterministic template. Human approval before publication. |
| Notifications | In-app only | Email/SMS channels are planned through the notification abstraction. |
| Incidents, users, teams | Fictional demo data | All people and incidents in the demo are invented. |
Data sources
Every chart, map and metric in CLIMATIQ shows which of these it came from and when it was last updated.
- Open-Meteo Forecast APIexternal apiactive
Numerical weather prediction guidance (best-match models). Used as an input to the CLIMATIQ baseline, not as an official forecast.
CC BY 4.0 (data); free API for non-commercial use · Weather data by Open-Meteo.com · docs
- Open-Meteo Historical Weather API (ERA5 reanalysis)external apiactive
Gridded reanalysis — a model-based reconstruction of past weather, not station observations. Lags real time by several days.
CC BY 4.0 (data); ERA5 by Copernicus Climate Change Service · Weather data by Open-Meteo.com · Contains modified Copernicus Climate Change Service information · docs
- Open-Meteo Previous Runs API (archived NWP as issued)external apiactive
Forecasts exactly as they were issued N days before the target day. Used so that the historical replay hindcast only uses information that was available at the time.
CC BY 4.0 (data) · Weather data by Open-Meteo.com · docs
- India Meteorological Department (IMD)officialnot configured
No openly permitted machine-readable API was available for this prototype. The adapter is implemented as "not configured"; official warnings are only shown when retrieved from a verifiable IMD source.
Per IMD data policy · India Meteorological Department · docs
- CLIMATIQ baseline-v1 modelmodelactive
Statistical blend of NWP guidance, persistence and a 5-year reference climatology.
Project code · CLIMATIQ model output (decision support, not an official forecast)
- CLIMATIQ demo simulatorsimulatedactive
Deterministic synthetic data used for demo stations and as an offline fallback. Always labelled SIMULATED.
Synthetic · Simulated demo data — not real measurements
- CLIMATIQ IoT ingestion gatewayiotactive
HTTP/REST ingestion for future IoT stations. No physical hardware is deployed in this prototype.
Project code · Registered IoT weather stations
Last successful ingestion: 3 Oct, 08:29 pm IST (station).
Forecasting method — baseline-v1
For every state and every district of the pilot states, CLIMATIQ forecasts the daily maximum temperature for the next 7 days by blending three ingredients:
- Reference climate — the average maximum for that date over 5 recent years of ERA5 reanalysis (±7-day window).
- Persistence — how far above or below that reference the last three days were; the anomaly fades with a 3-day time scale.
- Numerical weather guidance — Open-Meteo's forecast, weighted 85 % on day 1 falling to 55 % by day 7.
prediction(h) = w_h · NWP(h) + (1 − w_h) · [normal(d+h) + anomaly₀ · e^(−h/3)] σ(h) = 0.9 + 0.3·h + 0.25·min(|NWP(h) − persistence(h)|, 6) (+0.8 °C without NWP) range = prediction ± 1.28·σ (nominal 80 % band — heuristic, not calibrated)
Historical replay. To show the system on a real event, CLIMATIQ replays the late-May 2024 North-India heatwave: the forecast is made as of 26 May 2024 using only information available then — ERA5 history up to that day and the weather-model forecasts as they were issued (Open-Meteo Previous Runs API). It is then compared with what ERA5 later recorded. The replay is history, never a current warning.
The model is registered behind a versioned interface; every forecast run stores its model version, inputs and parameters so future machine-learning models can be compared fairly.
Severity classification
IMD declares a heat wave when a station's maximum reaches at least 40 °C in the plains, 37 °C on the coast or 30 °C in the hills and is 4.5–6.4 °C above normal (severe: more than 6.4 °C), or when it reaches 45 °C (severe: 47 °C) in the plains; IMD also requires two stations over two consecutive days (IMD FAQ on Heat Wave). CLIMATIQ maps these criteria onto four levels. Only the High and Extreme rules mirror IMD text; Moderate is a CLIMATIQ band. Thresholds are configurable by administrators (audited).
| Zone | Level | Rule | Mirrors IMD? |
|---|---|---|---|
| plains | Extreme | IMD severe-heatwave criterion: Tmax ≥ 40 °C and departure from normal ≥ 6.5 °C, or actual Tmax ≥ 47 °C. Departures use CLIMATIQ's 5-year reference climatology, not IMD's official normals. | Yes (approximation) |
| plains | High | IMD heatwave criterion: Tmax ≥ 40 °C and departure from normal 4.5–6.4 °C, or actual Tmax ≥ 45 °C. Departures use CLIMATIQ's 5-year reference climatology. | Yes (approximation) |
| plains | Moderate | CLIMATIQ-configured "approaching heatwave" class: Tmax ≥ 40 °C and departure ≥ 2.5 °C. Not an IMD category. | No — CLIMATIQ |
| plains | Low | Below all CLIMATIQ heat-risk thresholds. Not an IMD category. | No — CLIMATIQ |
| coastal | Extreme | IMD severe-heatwave criterion: Tmax ≥ 37 °C and departure from normal ≥ 6.5 °C. Departures use CLIMATIQ's 5-year reference climatology, not IMD's official normals. | Yes (approximation) |
| coastal | High | IMD heatwave criterion: Tmax ≥ 37 °C and departure from normal 4.5–6.4 °C. Departures use CLIMATIQ's 5-year reference climatology. | Yes (approximation) |
| coastal | Moderate | CLIMATIQ-configured "approaching heatwave" class: Tmax ≥ 37 °C and departure ≥ 2.5 °C. Not an IMD category. | No — CLIMATIQ |
| coastal | Low | Below all CLIMATIQ heat-risk thresholds. Not an IMD category. | No — CLIMATIQ |
| hilly | Extreme | IMD severe-heatwave criterion: Tmax ≥ 30 °C and departure from normal ≥ 6.5 °C. Departures use CLIMATIQ's 5-year reference climatology, not IMD's official normals. | Yes (approximation) |
| hilly | High | IMD heatwave criterion: Tmax ≥ 30 °C and departure from normal 4.5–6.4 °C. Departures use CLIMATIQ's 5-year reference climatology. | Yes (approximation) |
| hilly | Moderate | CLIMATIQ-configured "approaching heatwave" class: Tmax ≥ 30 °C and departure ≥ 2.5 °C. Not an IMD category. | No — CLIMATIQ |
| hilly | Low | Below all CLIMATIQ heat-risk thresholds. Not an IMD category. | No — CLIMATIQ |
Climate zones are a documented CLIMATIQ heuristic: hill states/UTs and Himachal Pradesh districts (except low-lying Una) use hill criteria; listed coastal districts (Odisha, Konkan) and coastal states/UTs use coastal criteria; everything else uses plains criteria. IMD does not publish this mapping.
Confidence & uncertainty
- Each forecast carries a range (nominal 80 % band) and a confidence label (low/medium/high) with a 0–1 score.
- The score is a heuristic from the spread σ — it grows with lead time and when weather-model guidance disagrees with persistence. It is not a calibrated probability.
- Forecasts are made at one point per district (or one point per state for non-pilot states). Cities display their district's forecast; CLIMATIQ never implies city-level precision.
- ERA5 reanalysis is a model reconstruction on a ~25 km grid; it tends to under-estimate local station extremes.
AI advisory generation
- Advisories are written from a validated forecast bundle (regions, temperatures, ranges, severity, confidence, contributing factors, sources). The model is instructed not to invent observations, sources, official warnings or certainty, and its output is checked against a schema.
- Source references are attached by the server, not by the AI. Provider, model, prompt version and generation time are stored with every advisory.
- If the AI provider is unavailable or no key is configured, a deterministic template writes the advisory and the fallback is recorded — the platform never blocks on AI.
- Every advisory starts as a draft and must be approved by an authorised official before it is published. Advisories are tailored for government, disaster-management teams, field teams and the public.
Automated alerts
After each forecast run, alerts are raised for regions whose forecast meets configurable rules (minimum severity, minimum confidence, maximum lead time). Duplicates are prevented per region, day and severity; a cooldown stops re-alerting after resolution; every alert is audited and delivered as an in-app notification to users responsible for that region. Alerts are always labelled as CLIMATIQ-generated — never as official warnings.
Forecast accuracy evaluation
Every forecast run is retained. When the truth becomes available, forecasts are verified against it (error = predicted − observed). For the historical replay, the truth is ERA5 reanalysis, not station observations, and it is a single event — so these numbers illustrate the method and are not a validated measure of skill.
| Lead time | Samples | Mean absolute error | Bias |
|---|---|---|---|
| Day 1 | 266 | 1.24 °C | +0.88 °C |
| Day 2 | 266 | 1.23 °C | +0.77 °C |
| Day 3 | 266 | 1.03 °C | +0.21 °C |
| Day 4 | 266 | 0.99 °C | -0.08 °C |
| Day 5 | 266 | 1.19 °C | -0.22 °C |
| Day 6 | 266 | 1.11 °C | +0.57 °C |
| Day 7 | 266 | 2.01 °C | +1.77 °C |
Detailed breakdowns by region, lead time and severity are in Climate analytics (sign-in required).
Demo & simulated data
- Demo weather stations and their observations are synthetic (no physical IoT hardware exists yet). They are always labelled SIMULATED and never mixed into real-data analytics.
- Users, teams, incidents, tasks and the response story are fictional. Demo records are flagged so the demo can be reset without touching real configuration.
- If real climate data cannot be retrieved, CLIMATIQ shows the last-updated time and a stale/missing notice — it never silently substitutes demo values.
Geography & resolution
Hierarchy India → state/UT → district → city. Boundaries come from geoBoundaries (state data from DataMeet, CC BY 2.5 IN; 2021 districts, ODbL), simplified for the web and depicting India's official claims; they are not authenticated by the Survey of India, which is the standard for official political maps of India. Cities are from GeoNames (CC BY 4.0). Districts are loaded for seven pilot states and UTs — Rajasthan, Uttar Pradesh, Maharashtra, Odisha, Telangana, Himachal Pradesh and Delhi — chosen to cover IMD's core heatwave zone and the plains, coastal and hill criteria. Some district boundaries predate later reorganisations.
Limitations
- Not an official warning; not validated for operational use; uncertainty bands are not calibrated.
- Departures use a 5-year reanalysis reference, not IMD's 30-year station normals; the IMD two-station/two-day declaration rule is not applied.
- Heat stress also depends on humidity, night temperatures, exposure and vulnerability; CLIMATIQ shows humidity and warm-night factors but does not model health impacts.
- Open-Meteo's free service has rate limits and no uptime guarantee; data freshness is shown everywhere.
Privacy & data handling
- Stored personal data is limited to account name, email, designation, role/region assignments and a bcrypt password hash. Demo accounts are fictional.
- Sessions use signed, http-only cookies. Actions that change data are written to an audit log with minimal before/after snapshots (no passwords or secrets).
- Only public weather data is sent to the AI provider — never personal data. Free AI tiers may use prompts to improve their services.
- Read notifications are deleted after 30 days; live forecast runs are kept for about 13 months for accuracy analysis. This prototype is not certified against any specific regulation.