Recorded presentation2.2
Arctic blizzard prediction by nonlinear statistical downscaling of GFS Reforecast model outputs
Zhen Zeng, University of British Columbia, Vancouver, BC, Canada; and W. W. Hsieh, W. R. Burrows, and A. Giles
Blizzards have been identified as hazardous weather events with the greatest impact on human activity in the Arctic, and as the greatest forecast problem there. For the purpose of improving Arctic blizzard forecasts, four nonlinear statistical forecast models, classification and regression tree (CART), Bayesian neural network (BNN), support vector regression (SVR) and Gaussian process (GP), in addition to linear regression (LR), were developed to forecast wind speed with 12, 24, 48 and 72 hr forecast lead times using daily reforecast data from the NCEP Global Forecasting System (GFS) nearest to two stations (Clyde River and Paulatuk) in the Northwest Territories and Nunavut regions in Canada during the 5 winters of 2000-2005. After scaling the forecasted wind speed to match the observed standard deviation, we compared the threat scores (TS) of calibrated forecasts from different models. The main findings were: (i) The nonlinear models provided better wind speed forecasts than the LR model. (ii) Among all nonlinear models, the SVR model gave the best cross-validated forecasts. (iii) The GP model presented the best performance on the blizzard forecasts at short forecast lead times (12 and 24 hr), while the SVR model surpassed others at longer lead times (48 and 72 hrs).
Session 2, Applications of Artificial Intelligence—II
Tuesday, 13 January 2009, 8:30 AM-9:45 AM, Room 125A
Previous paper Next paper