Forecast Post-Processing with Object-Based Analogs: A Technique to Reduce Spatial Dimensionality and Preserve Local Covariance

Maria Frediani

School of Engineering - University of Connecticut

Monday, Oct 24, 2016, 11:00 am
DSRC Room 2A305

Abstract

Analogue forecasting is a post-processing technique, in which a statistical forecast is derived from past prognostic states. Essentially, an analog set is sought according to similarities between the current and past forecasts, and a statistical forecast is derived from the corresponding observed states. Traditionally, analog sets are selected based on model quantities over a grid point, or point-values interpolated to a meteorological station geographical location.

This presentation describes a new method to seek analogs based on forecast objects, i.e., features with distinguishable shapes in a forecast field. Object-based analogs carry local spatial covariance, such that the best matches represent similar local dynamics. In addition, it reduces spatial dimensionality and allows for advanced algorithms to produce statistical predictions (e.g., artificial intelligence).

Forecast objects are created using a combination of image processing techniques that define objects based on spatially localized patterns. Although other techniques to select and match objects have been used in object-based forecast verification, they are inadequate to verify continuous variables, such as wind speed. To build the analog set, the predicted and observed quantity must exist, but the current object-based verification techniques cannot always find and match corresponding objects. Thus, a region-based technique for object matching is presented as a step in the Object-Analog methodology, and may be used on its own as a forecast verification technique.

The Object-Analog technique is validated using 10-m AGL wind speed forecasts and reanalysis, motivated by frequent storm-induced power outages in the Northeast U.S. Results from cross-validation are analyzed through statistical metrics comparing the method to a grid-based ensemble, and the ensemble mean against the raw forecast.