Title: Long-Range Forecasts Using Data Clustering and Information Theory
Abstract:
Even though forecasting the weather beyond about two weeks is not
possible, certain climate processes (involving, e.g., the large-scale
circulation in the Earth's oceans) are predictable up to a decade in
advance. These so-called climate regimes can influence regions as
large as the West Coast of North America over several years, and
therefore developing models to predict them is a problem of wide
practical impact. An additional central issue is to quantify
objectively the errors and biases that are invariably associated with
these models.
In this talk we discuss methods based on data clustering and
information theory to build and assess probabilistic models for long-
range regime forecasts. With reference to a simple ocean simulation
mimicking the Gulf Stream in the Atlantic (or the Kuroshio Current in
the North Pacific) we demonstrate that details of the initial state
are not needed in order to make skillful long-range predictions,
provided that an appropriate coarse-grained partitioning of the set
of possible initial conditions is employed. Here, that partitioning
is constructed empirically using running-average coarse graining and
K-means clustering of observed data, and optimized by means of
relative-entropy measures. We apply the same tools in a related
formalism for quantifying errors in imperfect climate models.
Together, these techniques provide a framework for measuring
predictive skill and model error in a manner that is invariant under
general transformations of the prediction observables.