Forecasting Monthly Rainfall in Australia

Error message

Deprecated function: Array and string offset access syntax with curly braces is deprecated in include_once() (line 20 of /usr/www/users/neurodim/neurosolutions/apps/includes/

Dr. John Abbot has been a scientific researcher for over 30-years both in academia and industrial settings. Currently, Dr. Abbot is a Professorial Research Fellow at Central Queensland University in Noosa, Australia. The majority of his projects are related to environmental issues including his current project for forecasting medium and long-term monthly rainfall in Australia. Dr. Abbot holds numerous degrees including a PhD in Chemistry from McGill University in Canada. Other degrees include Applied Computing, Technology Management and Applied Finance.

Australia is a continent that experiences severe droughts over extensive regions for many years, followed by heavy rains and severe flooding. Back in December 2010 and January 2011, there was very heavy rainfall in Queensland after nearly 10 years of comparative drought. This heavy rainfall caused severe flooding with most of the state of Queensland declared a disaster zone causing 40 deaths, billions of dollars in property damage, and extensive disruption of the mining industry. The Australian Bureau of Meteorology uses simple statistical models and complex physical models called General Circulation Models (GCMs). Unfortunately, this onset of heavy rainfall was not predicted months ahead by these conventional forecasting methods.

After working with other neural network software for a while, Dr. Abbot tried NeuroSolutions version 6 in 2012 and was able to generate a better forecasting model than both the competitor’s software as well as the Bureau’s best GCMs (see J. Abbot & J. Marohasy, 2014. Input selection and optimization for monthly rainfall forecasting in Queensland, Australia, using artificial neural networks, Atmospheric Research, Vol. 138, Pgs. 166–178).

Recently, Dr. Abbot moved his research to NeuroDimensions’ next-generation predictive data analytics software called NeuroSolutions Infinity. With it, he improved his previous results in NeuroSolutions by 10% (based on the MAE and RMSE) and 24% compared to the other neural network software he tried. These comparisons were made between the forecast result and the observational rainfall measurements over a ten year period. In addition, the neural network model results were compared against other rainfall forecasting models (GCMs) used by the Bureau of Meteorology.

The following table is representative of some of the results generated for forecasting monthly rainfall for locations in southeast Queensland over 80 months.

Table 1. NeuroSolutions Infinity neural network model monthly rainfall forecast for Harrisville, Queensland

Lead Time (months) Correlation RMSE (mm) MAE (mm)
1 0.78 37.4 27.7
3 0.77 38.3 28.9
6 0.76 39.0 30.2
9 0.82 39.0 29.0
12 0.76 39.8 30.4
18 0.75 40.0 29.7

Compare these results for the same period produced by the Predictive Ocean Atmosphere Model for Australia (POAMA) general circulation model, shown in in Table 2.

Table 2. POAMA monthly rainfall forecast for Harrisville, Queensland.

Lead Time (months) Correlation RMSE (mm) MAE (mm)
1 0.52 53.4 36.1
3 0.49 53.6 36.5
6 0.48 54.1 36.9
9 0.46 54.6 38.4
climatology 0.48 53.5 36.6

The neural network model generated by NeuroSolutions Infinity consistently generated better rainfall forecasts with higher correlation, lower RMSE and MAE statistics.

It is also important to visualize the forecast over time. The first graph shows the forecast monthly rainfall from the GCM for a 6 month lead (blue line). The red line shows the observed rainfall. If the model was perfect, the blue line would overlay the red line. In this case, the forecast blue line is not much different to taking the long term average rainfall for each month. Importantly, it does not give any warning of the very high peak at month 73 which corresponded to the floods of 2010/2011.

Figure 1. Monthly rainfall forecast results from the POAMA general circulation model with 6 months lead.

Now compare the rainfall forecast output from NeuroSolutions Infinity neural network model over the same test period. There is more evidence of modulation of the rainfall to better correspond with the observed measurements. In particular, for the forecast peak period at month 73 (December 2010), the neural network model provided a warning of impending heavy rainfall 6 months in advance!

Figure 2. Monthly rainfall forecast results from the NeuroSolutions Infinity neural network model with 6 months lead.

Dr. Abbot attended the NeuroDimension Online Neural Network Course in 2012 and started using NeuroSolutions soon after. Since that time, he has upgraded to NeuroSolutions Infinity using 8 processors on a single computer. Dr. Abbot fed over 100 input variables with approximately 1400 samples of data through NeuroSolutions Infinity which intelligently searched over 8600 input variations and over 4100 neural network models before reaching the current “Best” solution. The current “Best Model” is a Probabilistic Neural Network (PNN) with inputs including rainfall, local temperatures, and broad scale climate indices and a single output with the forecasted monthly rainfall.

Improving official forecasting methods will greatly benefit multiple industries including agriculture, mining and also water resource managers including dam operators. One of the largest potential beneficiaries of improved weather forecast is insurers and reinsurance firms. According to reports from the US, crop losses due to bad weather reached $10.7 billion for the 12 months to August 2012 with Australia’s insurer QBE’s losses on crop insurance as high as $270 million1 USD. Dr. Abbot says, “More skillful medium- and long-term weather and climate forecast with a particular focus on extreme events could significantly reduce the risk for the insured and insurer.”

When asked about what Dr. Abbot liked best about NeuroSolutions Infinity he stated the “user interface made [NeuroSolutions] Infinity very easy to use.” He also went on to say “… when introducing new people to running neural networks and becoming quickly productive, the [NeuroSolutions] Infinity interface is very well designed.”

Finally, when asked about why Dr. Abbot’s research group selected NeuroSolutions Infinity over other competitive products he stated that “[NeuroSolutions] Infinity is very good because of the automation in optimization of [neural] networks and inputs.” And that “it took us about a week to arrive at the result with [NeuroSolutions] NS 6, whereas the optimization with [NeuroSolutions] Infinity can be completed in about 8 hours.” Lastly, Dr. Abbot stated that the “best results to date are with [NeuroSolutions] Infinity.”

We would like to thank Dr. Abbot for the great insight on how he and his research group are striving to improve many industries in Australia with these improved forecasts and we look forward to hearing more from Dr. Abbot in the future!

Looking for predictive data analytics and neural network software to solve your own problem? Try NeuroSolutions Infinity free for 14-days! NeuroSolutions Infinity is the easiest, most powerful neural network software of the NeuroSolutions family. It streamlines the data mining process by automatically cleaning and preprocessing your data. Then it uses distributed computing, advanced neural networks, and artificial intelligence (AI) to model your data.