Thursday, 15 January 2004: 3:45 PM
Numerical forecast accuracy over the Northeast Pacific: Model intercomparison and cases of major failures
Short-term (up to 48 hours) predictions by numerical forecast models of cyclone position and central pressure can be in error as much as 100's of kilometers and 10's of mb, respectively, for strong North Pacific storms affecting the North American West Coast. These major numerical forecast failures still occur despite continued improvements
in numerical models and data assimilation systems. The freqency and intensity of sea level pressure (SLP) errors at buoy and coastal locations are documented by comparing 24-h and 48-h forecasts by several operational forecast models to observations for the 2002-2003 winter season. Results from the National Centers for Environmental Prediction (NCEP) Eta and Global Forecast System (GFS) models demonstrate that large forecast errors (absolute SLP errors greater than about 6 mb) occurred ~15 times this past winter and extremely large errors (over 10 mb) occurred 4 times. The Eta model had larger mean absolute SLP errors and standard deviations than the GFS model at most observation stations. Most of the 15 events with large forecast errors mentioned above were associated with low pressure systems. The
48-h forecast average position error and mean absolute central pressure error was larger for the Eta model than the GFS model for the 48-h forecast lead time (375 km and 9.3 mb for the Eta vs. 340 km and 5.3 mb for the GFS). However, the Eta and GFS 24-h forecast errors of cyclone position and intensity were very close (180 km and 4 mb). Further comparison of forecasts of Northeast Pacific cyclones by
numerical models from other operational centers are planned.
In addition, the role of initial condition errors in contributing to forecast errors are examined for cases with extremely large forecast errors. One such event occurred 7-8 February 2002. The storm was a compact, powerful cyclone that struck western Oregon with strong winds and produced extensive damage. The 24-h numerical forecasts from five operational models for this event were poor and a comparison of the initial conditions of the five models revealed large differences. The more accurate forecasts had sharper, more intense features in the initial state, whereas the less accurate forecasts started with initial conditions depicting a weak or incorrectly placed incipient cyclone. The role of initial condition errors in other poorly
forecast cyclones will be summarized and compared to the 7-8 February 2002 storm.