Teus — River and Flood Forecasting · Hasan Hüseyin Yurdagül

Teus forecasts hourly river discharge and evaluates every model on the rare flood events that average scores hide. It is a TÜBİTAK 1832 R&D project that I co-lead with a hydraulic engineer; I own the data, modeling, evaluation, and serving work.

24-hour seq2seq forecast closely tracking observed discharge as station D22A114 rises from about 5 to 72 cubic metres per second — Station D22A114, July 2023: a 24-hour seq2seq forecast (orange) against observed discharge (blue) as the river rose from about 5 to 72 m³/s in four hours. A selected event, forecast with observed rainfall as input. Labels are Turkish: *Debi* = discharge, *Tarih* = date.

The problem

Five years of hourly discharge and rainfall at station D22A114, with a handful of flood events marked in red — Five years at D22A114: discharge (dark blue), rainfall (light blue, inverted), and flood events (red). The events that matter are a tiny share of the record.

Floods are a small fraction of each station’s hourly record, which creates three traps:

A model that repeats normal conditions scores well on the full series and still misses every flood.
Flashy basins go from baseflow to peak within hours, so discharge history alone gives little warning; rainfall has to be an input.
Ordinary MAPE explodes near zero flow. Teus reports thresholded MAPE, scored only on hours above a station’s discharge-quantile threshold.

What it does

Compares classical and neural models, from Ridge baselines to a seq2seq LSTM with attention, on discharge and basin-weighted rainfall.
Scores peak-event windows with expanding-window backtests, separately from ordinary flow.
Writes every run’s predictions, metrics, plots, fitted model, and manifest to an inspectable report.
Serves checkpoints that pass loading and input-contract checks through a FastAPI forecast service.

Historical experiments covered 24–72-hour horizons. The current generation targets 1–12-hour forecasts across eight stations.

Teus run report for a Ridge baseline: run settings, a per-event metrics table, and RMSE and thresholded-MAPE charts — A run report for a Ridge baseline: 168 hours of history, a 72-hour horizon, three event backtests. Every run produces one, so weak baselines stay as visible as strong models.

Results and limits

Selected events show both sides of the model. At D22A114 it matched a sharp peak closely. At D08A084, a small and flashy basin, it caught the timing but not the magnitude:

Forecast at station D08A084 that rises at the right time but peaks near 5 cubic metres per second against an observed peak near 20 — Station D08A084, April 2021: the right timing, but about 5 m³/s predicted against an observed peak near 20 m³/s.

These are conditional experiments. Historical runs use observed future rainfall rather than archived weather forecasts, and flood thresholds are computed on the full series. No station yet passes the strict high-flow gate: at least 24 distinct targets, P99-event MAPE under 10%, and finite, non-negative predictions. The evaluation interval has also been reused for diagnostics, so the next accuracy claim needs a fresh, locked holdout.

Checkpoint serving works. An operational warning system still needs domain review and authorized use.