Comparison of techniques for the calibration of coupled model forecasts of Murray Darling Basin seasonal mean rainfall
Ensemble forecasts of South Eastern Australian rainfall from POAMA 1.5, a coupled oceanatmosphere dynamical model based seasonal prediction system run experimentally at the Bureau of Meteorology, tend to be under dispersed leading to overconfident probability forecasts. The poor reliability of seasonal forecasts based on dynamical coupled models is a barrier to their adoption as official outlooks by the Bureau of Meteorology. One approach to correcting this problem is model calibration, in which the probability distribution produced by the model is adjusted in light of available information about its past performance. Several distinct methods for calibrating seasonal rainfall forecasts for South Eastern Australia derived from the POAMA 1.5 ensemble are compared for accuracy and reliability in order to assess their suitability for application to real-time seasonal forecasts. The calibration methods investigated were: a variance inflation method (IOV); a Bayesian joint probability (BJP) calibration technique; and a singular vector regression technique (SVD) based on co-varying patterns of model and observed rainfall. Calibration was carried out for model grid points in the Murray Darling region. Assessment was carried out using a mix of standard skill scores widely used in operational forecasting. It was found that the BJP method resulted in the best correction to forecast reliability while IOV improved reliability only modestly and the SVD scheme had a negative impact on reliability. Further study of the application of these methods to real-time forecasts is recommended.
Collection(s) and Series: CAWCR technical report- No. 40
Format: Digital (Free), Hard copy (ill., charts)
ISBN (or other code): 978-1-921826-58-0