92nd American Meteorological Society Annual Meeting (January 22-26, 2012)

Wednesday, 25 January 2012: 5:15 PM
Calibrating Two Sophisticated Distributed Operational Hydrological Prediction Models for Romania: Final Results
Room 352 (New Orleans Convention Center )
John N. McHenry, Baron Advanced Meteorological Systems, LLC, Raleigh, NC; and T. K. Burnet, C. J. Coats Jr., S. Datta, R. E. Imhoff, D. Gochis, and C. D. Peters-Lidard

With the completion of the calibration of 3 complementary operational hydrological models, the modeling implementation phase of Romania's Destructive Water Abatement Program (DESWAT) is drawing toward a close. These include a classic lumped model approach together with sophisticated semi-distributed and fully-distributed modeling systems. All models are now running routinely at the Romanian national hydrological forecast center (INHGA), making use of both LDAS and FCST run components. The LDAS updates continuously, using precipitation observations from the network of Romanian weather radars combined with other forcing variables derived from operational Romanian NWP. Model deployment spans the entire country, which includes eleven major basins comprising about 230,000km2 -- an area slightly larger than the state of Minnesota.

The calibration effort for the distributed models (TOPLATS-SMPV, based on Peters-Lidard, et al., 1997, and Baron-LIS-NOAHv2, based on Gochis and Chen, 2004) took two-years to complete, including the significant efforts required to process, analyze, and prepare Romanian hydrological and meteorological observations, along with the development of formal protocols for using this data to calibrate the models during the six year period of record. This period included the major flood year of 2005.

Whereas TOPLATS-SMPV is deployed across 118 head-water catchments whose sizes average about 300km2, the Baron LIS_NOAHv2 model is deployed across 11 major basins, with single model instantiations running each major basin separately. Major basins average about 20,000km2 of modeled land-surface and stream-networks. The Baron LIS_NOAHv2 model simulates these integrated networks as unified entities including an average of more than 70 lakes and/or managed reservoirs per major basin. Thus, the calibration protocols had to respect the computationally intensive requirements for running multiple model realizations at a nominal model grid-spacing of 100-meters for both models.

The calibration effort was broken into a data-preparation phase followed by a two-phase set of calibration/regionalizations. The data preparation phase included the screening and merger of Romanian surface precipitation and temperature observations with ECMWF-ERA40 re-analysis data. This was subsequently downscaled to produce six-year continuous half-hourly input forcing meteorology at 1KM resolution to drive the land-surface components of the two DM's. It also included the screening and processing of more than 400 hydrological observations (initially collected using manual methods) to provide seemless 24-hourly average observed total and base-flow time-series.

Phase 1 of the calibration/regionalization effort involved the selection of six average headwater catchments and two regionalization-validation catchments for each of the two DMs. Headwater size catchments had to be utilized for calibrating both models because of the very short computational time-steps required by the explicit diffusive-wave-style overland and channel flow sub-models within the Baron LIS_NOAHv2.

Following Phase 1, the regionalization scheme was used to deploy an initially-calibrated operational version of both models across all catchments and major basins nationally. Phase 2 involved selection and calibration-validation of a second set of catchments, where a revised-updated set of regionalized parameters was developed by compositing/combining the calibration results from Phase 1 with Phase 2. Final results of the two-phase calibration-regionalization efforts are now fully deployed, completing the distributed model implementation for DESWAT.

Because of significant differences in their underlying conceptual-scientific bases, the calibration protocols had to be tailored for each of the two models. In general, however, both sets of protocols included both formal optimization and qualitative-subjective judgment in establishing final calibrated values for the 10 or so tunable parameters within each DM. While Phase 1 results showed significant improvement in all of the important statistical performance measures against a priori performance (RMSE, Correlation-Coefficient, Nash-Suttcliffe efficiency, and bias error), Phase 2 results further improved upon Phase 1. This provides high confidence in the overall protocols that were developed and utilized to calibrate the models.

To our knowledge, the above described implementation represents the largest operational calibrated distributed model deployment at a 100-meter nominal resolution running at any national center worldwide.This talk will describe the models used, the data-preparation phase, and the two-phased calibration process, including the calibration protocols and their final, implemented performance results.

Supplementary URL: