Numerical weather prediction is stochastic by nature. Twenty years ago, NCEP scientists decided that, with the limited computational power available then, they would have an ensemble of lower-resolution forecasts, rather than a single, deterministic prediction. For further details, please see this wonderful note by Steven Tracton and Eugenia Kalnay.
seamod.ro forecasts an ensemble of hydrodynamical models rather than a single model. Each member in the ensemble has different parameters:
– initial conditions are modified using realistic perturbations generated by WCE 1.0, separately for the basin and the coastal zones.
– internal model parameters are chosen randomly in a range of plausible values
– river outflow is also perturbed in a realistic range
– atmospheric forcing fields are perturbed using principal component analysis, in such a way that the perturbations are physically realistic, and their intensity correspond to the expected uncertainty. The latter is estimated by comparing forecasts from 2 different weather prediction centers.
The standard deviation of the ensemble is then used as a proxy for the expected forecast error. Because the ensemble has to be simulated entirely, uncertainty forecasts lag 1 day compared to the forecast itself.
Below, you can see yesterday’s 1-week forecast, as well as the associated forecast uncertainty, for sea surface elevation, surface current velocity, and temperature. Standard deviations being by definition positive, the surface current velocity uncertainty is always pointing north- and east-ward. As expected, please also observe how forecast uncertainty increases with lead time.
You can download or plot other uncertainty forecasts from our Thredds server (“std” files)!