A Distance-based Methodology for Comparing Longer-Term, Multi-Metric Model PerformancePrimary Author: Brekke, Levi
Additional Authors: Michael Anderson, Michael Dettinger, Edwin Maurer
Brekke*1, Levi D., Michael Anderson2, Michael Dettinger3, Edwin Maurer4
1U.S. Bureau of Reclamation, Technical Service Center, Denver Federal Center, Denver, CO, 80225
2CA Dept. of Water Resources, Div. of Flood Management, Sacramento, CA
3USGS and Scripps Institution of Oceanography, La Jolla, CA
4Santa Clara University, Civil Engineering Department, Santa Clara, CA
email: lbrekke@do.usbr.gov
A Distance-based Methodology for Comparing Longer-Term, Multi-Metric Model Performance
In an ongoing regional risk assessment, climate projection distribution functions are being developed for Northern California, representing an ensemble of IPCC AR4 projections produced by 17 coupled models simulating either SRES A2 or B1. It is questioned whether these distribution functions should built to reflect unequal model-weighting derived from relative model performance in the 20th Century Climate Experiment (20C3M).
To explore this question, 59 20C3M simulations from the same 17 models were evaluated statistically during 1950 1999. These statistics were then compared to reference conditions during the same period (NCEP/NCAR Reanalysis, Kaplan Extended SST v2). This comparison was made for a range of statistical metrics applied to different global variables, local variables, and teleconnections relevant to Northern California climate (i.e. local precipitation and surface air-temperature, North Pacific sea level pressure index, and Nino3 sea surface temperature index). Metric-specific differences were then aggregated using a distance scheme to reveal multi-metric model-to-reference similarity. Relative model weights were derived from similarity calculations and ultimately used in producing weighted estimates of climate projection distribution functions.
Presentation will provide an overview of methods, but focus mainly on results from 20c3m similarity analyses. Results show that although bias for a given metric can vary significantly among the models analyzed, the relative degrees of bias varies considerably depending on the variable and/or metric. Moreover, consideration of multiple metrics was found to significantly dampen the range of relative model weights.
|
