A Model-output-calibration-based Consensus Forecast Method for Tropical Cyclone Intensity

Forecasts of tropical cyclone (TC) intensity up to 120 h from seven operational models (three global models and four regional models) during three years (2010-2012) are analyzed in this study and a model-output-calibration-based consensus forecast method (MOCC) for tropical cyclone intensity is proposed. The three global models are from CMA, ECMWF, and NCEP, respectively. Two regional models are from CMA and the other two are from FNMOC. It is found that the model forecast errors are highly correlated to the initial state of the TC and its environmental conditions, including initial TC intensity, size, moving speed, environmental pressure, and the maximum potential intensity as determined statistically by sea surface temperature. Intensity forecasts by most of the models are too strong for weak TCs and too weak for strong TCs at short lead times. All the models tend to be too strong for large TCs and too weak for small TCs. The forecast errors are also found to be significantly correlated with initial intensity error of the model. A statistical calibration scheme for TC intensity forecasts of the models is then proposed based on the above analyses by using stepwise regression technique. Independent test demonstrates that the proposed calibration scheme is effective for all the studied models in improving their forecast skill of TC intensity. Two most important factors for calibration equations at short lead times (24h and 48h) are the initial error of model and the observed intensity change in 12 hours before the initial time. Initial TC size is a good predictor for the calibration of three global models at long lead times (72h, 96h and 120h). Two multi-model consensus methods are experimented, with one based directly on the TC intensity forecasts of the models and the other one based on the statistically calibrated model forecasts, namely the MOCC. Independent tests of MOCC show forecast skills > 10% for all the lead times over a statistical baseline, implying that the calibration is important in setting up a skillful multi-model consensus for TC intensity.