Techniques for Down-selecting Numerical Weather Prediction Ensembles

Open Access
- Author:
- Lee, Jared Armand
- Graduate Program:
- Meteorology
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- June 22, 2012
- Committee Members:
- Sue Ellen Haupt, Dissertation Advisor/Co-Advisor
Sue Ellen Haupt, Committee Chair/Co-Chair
George Spencer Young, Committee Chair/Co-Chair
David R Stauffer, Committee Member
Marcelo Chamecki, Committee Member
Derrick J Lampkin, Committee Member
Leonard Joel Peltier, Special Member - Keywords:
- numerical weather prediction/forecasting
ensembles
statistical modeling
model evaluation/performance
down-selection - Abstract:
- Ensembles of numerical weather prediction (NWP) models are valuable forecast- ing tools for a number of applications. Because the spread in the ensemble predictions is correlated to forecast uncertainty, NWP ensembles quantify the uncertainty of the prediction. Ideally NWP ensemble spread should accurately represent uncertainty in the low-level mean wind for forecasting applications such as wind energy and atmospheric transport and dispersion. To adequately sample the probability density function (PDF) of the forecast atmospheric state, all relevant sources of uncertainty ought to be ac- counted for, including model error. Multi-physics ensembles can be used to account for the kind of model error that arises from imperfect knowledge of physical processes. This study presents the first known objective methodology to guide users in choosing which combinations of physics parameterizations to include in an NWP multi-physics ensemble. For this work two NWP multi-physics ensembles are built using the Advanced Research Weather Research and Forecasting (ARW-WRF) model. For Part 1, I construct a 24-member ensemble for 48-h forecast periods over a summer season. For Parts 2 and 3, I build a 42-member ensemble for 48-h forecast periods every fifth day for winter, summer, and a transition season. Verification for Part 1 focuses on surface temperature and wind components at forecast lead times from 12-48 h. Both surface and low-level (925, 850, and 700 hPa) temperature and wind components are used in Parts 2 and 3. In Part 1, I introduce the down-selection methodology using principal component analysis (PCA) over a summer season. PCA yields a down-selected 14-member ensemble that has similar performance to the full 24-member ensemble, as measured by root-mean square error (RMSE), and continuous ranked probability score (CRPS). In Part 2, I compare the performance of a refined PCA technique with two additional down-selection techniques, K -means cluster analysis (KCA) and hierarchical cluster analysis (HCA) over a winter season. All three techniques yield down-selected ensembles of 10-11 members that perform equally well as the full 42-member ensemble, using RMSE, CRPS, and the normalized reliability index (RI). In Part 3, I explore the dependence and performance of the HCA clustering and down-selection results on seasonality and length of training period, and also examine spread-skill distributions of the full and subset ensembles. For the full and down-selected ensembles throughout this work, the ensemble PDF is also statistically dressed, or calibrated, using Bayesian model averaging (BMA). Using statistical down-selection techniques such as those outlined here allows for a small multi-physics ensemble with a few intelligently chosen members to represent model error and the forecast PDF just as well as a multi-physics ensemble that is four times as large. This approach saves computing resources, allowing NWP ensemble modelers to use those resources for other important aspects, such as increasing resolution or representing additional sources of uncertainty.