On The Limits Of Predictability Of Day To Day Weather Forecasts For Melbourne Australia

The authors have recently completed a piece of work exploring trends in the skill of weather prediction at lead times of 1 to 14 days for Melbourne, Australia (QJRMS, 2015). Official Australian Bureau of Meteorology forecasts were used to establish these trends at shorter lead times - out to Day-7. The system that was used to establish these trends at longer lead times - out to Day-14 - was, in part, based upon an algorithm that statistically interpreted the GFS NWP model output to generate local weather forecasts. More recently, the application of other NWP models towards determining predictability limits has also been explored. To this end, the authors presented preliminary results to the 2015 American Meteorological Society Annual Meeting about what had been achieved using a statistical interpretation of the output (over a six-month period) of the ECMWF monthly control model (which generates predictions out to Day-32). Since then, further sets of GFS and ECMWF model output data have been collected. These have been combined optimally (in the context of a multi-model ensemble) in order to investigate how that process (of combining) enhances predictive skill. Preliminary results of this exercise were presented to the 2015 Australian Meteorology and Oceanography Annual Conference and it is the purpose of the current paper to utilise the additional data collected since then to formulate conclusions with a higher level of confidence than hitherto. These conclusions are that the process of combining does, indeed, enhance predictive skill and that there is some capability, albeit limited, exhibited by the day to day forecasts based upon the output of the ECMWF control model beyond Day-14. To illustrate this latter point, the first graphic shows the accumulated percent variance of the observed rainfall explained by the Day-15 forecasts, whilst the second graphic shows the accumulated percent variance of the observed maximum temperature explained by the Day-15 forecasts. Initially, when the data bases were small (each extending over only a few months) the accumulated percent variances fluctuated sharply each time a new data element was added (at times dipping below zero). Later, as the size of the data bases grew, the fluctuations diminished. After nearly a year, the percent variance of the observations explained by the forecasts is shown, in both cases, to have converged to positive, albeit very small, values.