Object-oriented clustering analysis of CAPS convective scale ensemble forecasts for the NOAA Hazardous Weather Testbed Spring Experiment: A first step toward optimal ensemble configuration for convective scale probabilistic forecasting

This study seeks a systematic understanding of the impacts and importance of sampling different sources of uncertainties using the convection-allowing ensemble run by Center for Analysis and Prediction of Storms (CAPS) during the Hazardous Weather Testbed Spring Experiment. The goal of the study is to aid optimal ensemble configuration by inferring issues that prevent more effective communication of uncertainty and probabilistic forecast guidance.

Recent advances in computational resources allowed production of an experimental real-time ensemble to be run for 5 weeks each spring over a near-CONUS domain at a 4-km convection-allowing resolution. The ensemble contains perturbations to initial and lateral boundary conditions (IC/LBC), model dynamics, and model physics. The data set provides a unique opportunity to understand how to optimize such an ensemble for explicit prediction of the timing, location and mode of severe storms.

A renovated Hierarchical Cluster Analysis (HCA) was conducted. Instead of using Euclidean distance (ED) to measure dissimilarity for precipitation forecasts a new object-oriented measure, called Fuzzy Object-oriented Threat Score (fuzzy OTS), was defined and adopted in the HCA. Advantages of Fuzzy OTS, relative to a neighborhood based ED include an ability to identify similar features in slightly different locations and a reduced sensitivity to overall forecast precipitation amount.

Precipitation forecasts are found to cluster based on model dynamical core and microphysics scheme at 3 hr lead time (valid 03 UTC), based on model, planetary boundary layer (PBL) scheme and IC/LBC perturbation at 12 hr lead time (valid 12 UTC), and based on model and PBL scheme at 24 hr lead time (valid 00 UTC).

Traditional formulation of HCA is also used for other variables. It is shown that 10 m wind speed forecasts cluster primarily based on PBL scheme, with secondary clustering based on model and IC/LBC perturbation, at early lead times. At later lead times 10m wind speed clusters are primarily based on IC/LBC perturbation. 500-hPa temperature forecasts cluster based on IC/LBC perturbation at all lead times. Attempts will be made to explain the observed behaviors.