CQUniversity researchers - including one who honed his skills through share market trading – have been able to forecast seasonal rainfall more accurately than the Met Bureau by using artificial intelligence for pattern analysis…
A journal, Advances in Atmospheric Sciences, is about to publish a new paper by CQUniversity researchers Dr John Abbot and Dr Jennifer Marohasy. The paper shows the application of artificial intelligence to monthly and seasonal rainfall forecasting in Queensland, Australia, and compares their forecasts with forecasts by the Australian Bureau of Meteorology’s super computers.
In March 2009 the Australian government ordered two new supercomputers ostensibly to ensure Australia is at the forefront of international weather
The super computers run a general circulation model known as POAMA that is a mathematical representation of general atmospheric circulation patterns. In the West, attempts to improve rainfall forecasts from general circulation models have focused on improving our basic understanding of weather processes, most recently with a focus on the role of carbon dioxide as a greenhouse gas.
In other parts of the world (notably China, India and Iran), governments have also funded research into artificial neural networks for rainfall forecasting. This radically different method is based on pattern analysis accepting that there are patterns, for example short and longer-term cycles, evident in rainfall data. Neural networks, based on artificial intelligence, have the ability to consider large numbers of climate indices (eg. El Nino, Indian Ocean Dipole) and other inputs (eg. temperature, cosmic ray flux) simultaneously and make predictions independently of any understanding of, for example, the hydrological cycle.
Following the devastating floods of January 2011, with three-quarters of Queensland declared a disaster zone, CQ-based researchers Dr Abbot and Dr Marohasy combined their respective interests in artificial intelligence for pattern recognition with climate science to see if they could forecast the weather at least as well as the Australian Bureau of Meteorology.
In their first attempt at optimising their dynamic stand-alone time-delay recurrent neural network (TDRNN) they found that patterns within rainfall data alone could provide a forecast.
Their optimal model combined current and lagged rainfall, temperatures, Southern Oscillation Index, Pacific Decadal Oscillation and Nino 3.4 reflected in the highest Pearson correlation coefficient and lowest root mean squared error value (RMSE). This model was applied to 20 sites across Queensland generating monthly rainfall predictions three months in advance.
The Australian Bureau of Meteorology provided output data from POAMA thus enabling a direct comparison of the ability of the two models to forecast seasonal rainfall: the general circulation model (POAMA) developed from a theory of climate and run on a super computer with a large staff versus the neural network prototype based on pattern recognition theory and run off a laptop in a small office in CQUniversity.
Outputs, as monthly rainfall forecasts three months in advance for the period 1993 to 2009, were compared with observed rainfall data using time-series plots, RMSE and Pearson correlation coefficients. The comparison showed the prototype neural network achieved a lower RMSE for 16 of the 17 sites compared, meaning it gave a better forecast for 16 of the 17 sites.
Dr Abbot, who honed his skills with neural networks through share market trading, considers the prototype design for rainfall forecasting very preliminary, with potential for significant improvement and application for anywhere in eastern Australia with at least 100 years of high quality historical rainfall data.
The findings have significant application to everyone affected by the weather but particularly for agriculture and mining with scheduling of mine activities in Central Queensland significantly impacted by wet days.
