|Subject||Improvement of real-time forecast system on the BSISO (Ⅱ): Multi-model ensemble forecast (Korean)|
|Author||Dr. Haejung Kim||Date||2016.01.01|
The objectives of this study are to improve BSISO forecast skill by developing MME method and to provide more updated BSISO information by setting up a stable operation system.
Regarding the development of BSISO MME method, the Simple Composite Method (SCM), Simple Linear Regression Method (SLR), and Multiple Linear Regression Method (MLR) are used as linear methods, while the Genetic Algorithm (GA) is utilized as the non-linear method. Four operation models are used when applying the MME method. The number of available forecast samples was 46 days with a 151 day data window. Pattern correlation, Hit Rate, False Alarm Rate, and Bivariate Correlation methods were used to verify anomaly fields and BSISO indices. The MME performance in predicting the BSISO varies with the methodology. BSISO MME indices are able to be well predicted with a 1 week to 2 week forecast lead time. SCM has the best performance in predicting BSISO indices and its skill is comparable to the best performance of a single model. Calibration can help to improve the accuracy of OLR and U850 anomalies. Application of an optimal technique for each variable may be needed to produce more skillful BSISO indices.
For the BSISO operational work, the observation data is changed from NCEP Reanalysis 2 to NCEP Reanalysis 1. This change provides more stable and up-to-date monitoring and forecast information by cutting the gap between the date of the latest observation and today’s date from 7 days down to 3 days. In addition, BSISO monitoring indices are able to be downloaded from the APCC webpage, which is more user friendly. Taiwan CWB's forecast model has joined in the BSISO forecast activity as a participant and its forecast service has been available from May 1, 2015.