Hydrosphere
How far the river sits from its normal state for the season — flood pulses and drought lows both stress the corridor, and both move the dissolved-solids concentration estimated from flow.
Withheld by the Trust Core: the calibrated band is too wide to be actionable. The platform reports this rather than passing a number it cannot vouch for.
Stress basis: distance from the seasonal median flow (percentile): 12th percentile of 671 same-season days across 11 years.
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 13-feature input row from the training set. Furthest feature: Season (sin) (-1.42σ).
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)
Day-of-year phase
GloFAS discharge on the Hantan gorge reach
FAO reference ET₀, the basin's water loss
ERA5 daily mean 2 m temperature
ERA5 daily snowfall (delayed runoff)
GloFAS discharge on the neighbouring reach
Basin rain over the last 3 days
Basin rain over the last week
GloFAS discharge on the Imjin wetland reach
ERA5 daily precipitation (wash-out / runoff)
Day-of-year phase
Monthly rain, the soil-moisture proxy
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=0.100 (band ×1.00 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 net2,061 params | neural · recurrent | 29% | 0.141 | 51.14 |
| MLP feed-forward net2,149 params | neural · feed-forward | 33% | 0.143 | 52.54 |
| AR(7) ridge | statistical | 20% | 0.165 | 45.82 |
| Holt smoothing | statistical | 5% | 0.257 | 41.18 |
| Seasonal naive | baseline | 1% | 0.725 | 174.79 |
| Persistence | baseline | 13% | 0.176 | 44.08 |
* held-out test window, log-transformed units
Water-quality layer
Dissolved solids are estimated from discharge with a published concentration–discharge power law, using literature parameters for temperate headwater streams, and labelled as an estimate.
Provenance
Governed quantity: Next-day GloFAS river discharge on the Bukhan River reach at Gapyeong (5 km grid cell).
archive pulled 2026-10-02 09:52 UTC · live tail merged at request time · all sources publicly available · nothing on this page is simulated