home » about nssl » research divisions » forecast r&d » models and assimilation
Data Assimilation
One of the challenges of numerical weather prediction is creating an initial state, or "picture", of the atmosphere from which to start a model forecast. This initial state is created using as many of the available observations as possible. There are a number of challenging problems that arise in data assimilation for mesoscale and storm-scale numerical weather prediction. One problem is that the model resolution is higher than that of the available observations, such that the model predicts features that cannot always be observed. One approach to solving this problem is to run the model in a pre-forecast assimilation period, in which the available observations are used to constrain the model forecast toward the observations. This approach often uses an adjoint version of the forward model, and involves slowly varying the model initial state until the model forecast best matches the observations during the assimilation period. This initial state is then used to make a full model forecast out to a predefined number of hours or days. Another approach is to use our knowledge of the atmosphere to construct features within the model initial conditions that we know are present, but we cannot fully observe. Both approaches are being explored by scientists within the team.
A second problem is that some observations do not include direct information on the standard variables that are contained within the model equations. For example, a Doppler radar only measures the wind in the direction along a radar beam, and can provide no direct information on the component of the wind that is moving perpendicular to the radar beam. Yet the model uses information on the total wind field, which includes both of these wind components. Owing to these types of problems, data assimilation techniques are again used to incorporate non-standard data into numerical weather prediction models and evaluate their usefulness in improving forecasts.
