Technical details
This page explains where the data comes from, how it is prepared, and which limits shape the forecast before it reaches the site.
Forecast model
NO2 model v1
+5.6% vs. baseline
Horizon
24h
28 stations · 12 pollutants
Training window
2025-01-08 → 2025-12-31
358 training days
Test window
2026-01-01 → 2026-04-29
119 validation days
Current operational read
Cloudflare D1 + ingestion
The API is already reading from Cloudflare D1, although the last valid observation is still 28 Sept 2026, 02:00. The next point to review is the protected ingestion cycle, not database connectivity.
Environment
production
reported by the API
Observation source
Cloudflare D1
active read path
Latest valid observation
28 Sept 2026, 02:00
over 12h old
Remote ingest
cron ready
D1 yes · jobs yes
If you only want to interpret current conditions, start with the overview, the map, or the forecast. This page goes through the full process: data, preparation, validation, and limits.
Back to guideComunidad de Madrid: official 2025 and 2026 historical air-quality exports.
Cloudflare D1 serves the recent normalized observation signal, while official daily CSV exports remain the verified raw input and fallback source.
Official station metadata for municipality, address, zone, and coordinates.
Weather variables remain documented for future iterations, but they are not part of the published v1 model yet.
Flexible ingestion of official CSV files with pandas.
Wide-to-long normalization into station, pollutant, and hourly observations.
Consolidation of `.normalized.csv` snapshots with timestamp-level deduplication before the operational signal is published.
Conservative `risk_level` assignment and explicit removal of invalid values.
Current observation coverage: 28 stations and 12 active pollutants.
No synthetic stations, no invented coordinates, and no frontend-side forecasting.
Freshness and consistency checks gate `/api/latest`, `/api/summary`, and `/api/system-status`.
ML artifacts are published only after an honest temporal split and baseline comparison.
Temporal variables: hour, day of week, month, and weekend flag.
Station encoding to capture structural differences across official nodes.
NO2 lags at 1h, 3h, 6h, 24h, and 168h.
3h, 6h, and 24h rolling means computed per station without temporal leakage.
The current validation uses train 2025 / test 2026; an intermediate persisted validation window does not exist yet.
Training window
2025-01-08 → 2025-12-31
Test window
2026-01-01 → 2026-04-29
Real comparison between the selected predictor and the reference baselines on the test window.
persistence
n=78077
same_hour_yesterday
n=78077
hist_gradient_boosting_v1
n=78077
The public API still does not expose station-level error even though the artifact computes it.
Observations already come from Cloudflare D1, but freshness still depends on the protected ingest cycle completing on time.
Weather features are not part of the current v1 model.
Forecasts and metrics are still published as versioned artifacts, and weekly historical reporting is not persisted yet.