Atmosphere
Particulate load over the DMZ corridor, with the Fire Weather Index and satellite fire detections as the acute-hazard context. The fastest-moving stream the platform watches.
Inputs in-distribution, members in agreement, interval tight — the Trust Core stands behind this number.
Stress basis: PM2.5 vs the 35 µg/m³ 'bad' threshold: 9.3 µg/m³ against a midpoint of 35 µg/m³.
Ensemble spread
Six heterogeneous members — a GRU, an MLP, a ridge autoregression, Holt smoothing, a seasonal baseline and persistence — each vote on the next value. Their weighted scatter relative to the recent one-step error is the disagreement signal; wide scatter means the members have genuinely split.
Trust signals
Spread of members vs expected one-step error.
Standardised distance of today's 14-feature input row from the training set. Furthest feature: Fire Weather Index (+1.78σ).
The combined trust the Core assigns this output.
What drove this prediction
Input×Gradient attribution: the gradient of the neural members' forecast with respect to every input, multiplied by the input, summed over the window — so each named driver carries a sign and a magnitude. The AR member's coefficients fold into the target's own history.
The series' own recent values (autoregressive memory)
CAMS dust, the trans-boundary Yellow-Sea signal
CAMS daily mean PM10
CAMS daily mean NO₂, a combustion tracer
ERA5 daily mean relative humidity
Canadian FWI computed from ERA5
Day-of-year phase
ERA5 daily mean 2 m temperature
ERA5 daily precipitation (wash-out / runoff)
CAMS daily mean surface ozone
Day-of-year phase
ERA5 daily max VPD, the drought-stress driver
Column aerosol load from CAMS
ERA5 daily maximum 10 m wind (ventilation / dispersion)
This stream, audited
7-step path
The ensemble is rolled forward recursively; each horizon carries its own conformal quantile from the calibration window, scaled by the adaptive factor, so the band widens honestly with lead time. Drivers are held at their last value and only the season advances.
Rolling backtest, last 90 steps
Every point is a real one-step forecast made from the data available at the time, against what the feed then reported. Misses outside the band are marked. Adaptive conformal inference nudges α each step toward the 90 % target — currently α=0.140 (band ×0.85 vs static).
Model card
trained 2026-10-02A heterogeneous ensemble: two neural networks trained by backpropagation on the archive, two classical forecasters, and two baselines every member must beat. Weights are set by validation error, never by the test window.
| Member | Kind | Weight | Test MAE* | Vote |
|---|---|---|---|---|
| GRU recurrent net2,121 params | neural · recurrent | 25% | 0.359 | 9.18 |
| MLP feed-forward net2,289 params | neural · feed-forward | 21% | 0.366 | 8.23 |
| AR(7) ridge | statistical | 19% | 0.379 | 10.19 |
| Holt smoothing | statistical | 16% | 0.412 | 9.78 |
| Seasonal naive | baseline | 7% | 0.616 | 14.75 |
| Persistence | baseline | 12% | 0.402 | 7.23 |
* held-out test window, log-transformed units
Fire layer
The Canadian Fire Weather Index computed from ERA5 over the corridor, plus every VIIRS active-fire detection inside the watched box in the last seven days.
Provenance
Governed quantity: Next-day corridor-mean PM2.5 across the six sites.
archive pulled 2026-10-02 09:55 UTC · live tail merged at request time · all sources publicly available · nothing on this page is simulated