Aller au contenu principal
OceanData Consulting
Scoreboard · 48 h alerts

A wind or wave peak within 48 hours?

Every morning, the AI post-processing issues for each station the probability that an exceedance threshold is crossed within two days, with the time of the peak. What is shown here is today’s issue, checked the next day against the instrument.

Today’s issue, all stations

The threshold is the station’s upper decile over its training history: for tides it applies to the surge, never to the level. A station “without an issue” is not a calm station — its instrument delivered no observations, and the model refuses to issue without them.

48 h alerts — Figures not displayed

Measurements are read when the page opens, from the pipeline’s public repository; they are never frozen into this page. They could not be obtained here — network unavailable, file not matching the contract, or page read without JavaScript. Nothing is shown rather than a doubtful figure.

Check the public repository status

Réseau indisponible (scores.json)

What the alert is worth

Two measures, never mixed. The training holdout: detection and false alarms over a year of replayed issues, as published in gate.json at the last training. Production: the same indicators over the days actually lived since going live, day after day, without retouching.

Alert scores — Figures not displayed

Measurements are read when the page opens, from the pipeline’s public repository; they are never frozen into this page. They could not be obtained here — network unavailable, file not matching the contract, or page read without JavaScript. Nothing is shown rather than a doubtful figure.

Check the public repository status

Réseau indisponible (scores.json)

Your instrument

The same service on your sensor

Everything displayed on this page is produced from three inputs. If you have them, the service can be deployed on your measurement point, with the same daily publication and the same verdict.

01

An hourly history of the instrument

At least one year of continuous observations: the wind and wave stations of this scoreboard are evaluated rolling-origin over 360 days of issues, the tide gauges on an annual holdout. A shorter history yields a “degraded” verdict, displayed as such.

02

A near real-time observation feed

The model reads the latest measurements before each issue. Without them, it does not issue — that is the “no issue” state visible above, not a guessed figure.

03

A reference physical forecast

The model corrects an operational forecast, it does not replace it: here the wave and wind models served by Open-Meteo, for you your own model or a provider’s, chosen station by station at training.

04

A verdict before publication

The quality gate decides at training whether the alert can be published. After that, every issue is checked against observations and the scores accumulate in public — missed days included.

Protocols and history depths: those published per station in gate.json. Full method on the method page.

Contact

Do you have a deployed instrument?

Send us an extract of its history and the name of your reference forecast: within a few days we tell you whether a publishable verdict is attainable.