Data assimilation of local model error forecasts in a deterministic model

One of the most popular data assimilation techniques in use today are of the Kalman filter type, which provide an improved estimate of the state of a system up to the current time level, based on actual measurements. From a forecasting viewpoint, this corresponds to an updating of the initial conditions. The standard forecasting procedure is to then run the model uncorrected into the future, driven by predicted boundary and forcing conditions. The problem with this methodology is that the updated initial conditions quickly 'wash-out', thus, after a certain forecast horizon the model predictions are no better than from an initially uncorrected model. This study demonstrates that through the assimilation of error forecasts (in the present case made using so-called local models) entire model domains can be corrected for extended forecast horizons (i.e. long after updated initial conditions have become washed-out), thus demonstrating significant improvements over the conventional methodology. Some alternate uses of local models are also explored for the re-distribution of error forecasts over the entire model domain, which are then compared with more conventional Kalman filter type schemes. Copyright (C) 2002 John Wiley Sons, Ltd.