Several people have emailed me an article from the Australian Broadcasting Corporation, with claims that some startups can provided more accurate rainfall forecasts than the Bureau – using artificial intelligence (AI).  The Bureau’s counter claim, repeated in the same article, is but, AI is not good for long-term forecasts.

It is not news that the Australian Bureau of Meteorology is not very good at seasonal weather forecasting.

They mostly keep forecasting below average rainfall, or drought, and then we get another flood.   That is what happened last summer, for northeastern Australia – and it happens over and over with their forecasts for the Murray Darling Basin.

As I have written over and over, the Bureau don’t bother to benchmark how bad they are – though I have.  In a series of technical papers published in international climate science journals with John Abbot, beginning in 2012 and ending for me in 2017, we documented an alternative and better technique using artificial intelligence (AI).  In the first of these papers* we compared output from the Bureau’s simulation model with our AI-based statistical model for 17 locations in Queensland.

It was no small task getting the Bureau to provide the data allowing the comparison to be made – that was achieved in August 2011.   In the same meeting at the Bureau’s headquarters in Melbourne, I outlined to the then head of their long-range weather forecasting unit the possible benefits of work with Abbot and I, to further develop the technique.

We were the first to demonstrate the value of AI for rainfall forecasting in Australia.

I hope that the start-ups that are now focused on short term forecasts have more commercial success than we did.   AI has so much application for weather forecasting – short, medium term, seasonal and long range.

There is something about the history of my work with Abbot in the most recent issue of the IPA Review, CLICK HERE.   This article includes comment by me:

Locations along the east coast of Australia, including Cairns and Lismore, are very affected by changing sea surface temperatures and pressures across the South Pacific that have been measured since the late 1800s.

The official temperature database for Australia, known as The Australian Climate Observations Reference Network – Surface Air Temperature (ACORN-SAT) is used to generate an average Australian temperature, and this database only begins in 1910.

This record only begins in 1910 because many weather stations did not record temperature in what is known as a Stevenson screen (basically a white louvred box) until about 1910. Before 1910 the mercury thermometers used to measure maximum temperatures were not necessarily kept in a standard housing, and this could result in higher temperatures for the same weather.

A Stevenson screen did not become the official housing for the thermometers at the Bureau’s official weather station in Sydney until 1910. In Melbourne a Stevenson screen was not installed until 1908, and in Brisbane a Stevenson screen was installed earlier in 1896.

Darwin has the longest record, with an official temperature record from a mercury thermometer in a Stevenson screen starting in March 1894. This is all documented in the online archive for Darwin at the Bureau’s website, and I have found photographs of different shelters and other instruments in the Darwin public library including a photograph taken in January 1890 of a Stevenson screen in the post office’s yard.

Charles Todd is the person to thank for Darwin’s exceptionally long, continuous, and reliable early temperature record. He was an avid meteorologist, astronomer, and electrical engineer who oversaw the construction of the Overland Telegraph line connecting Darwin with Adelaide that was completed in 1870. That was the same year Todd became Australia’s first Postmaster-General.

After the completion of the Overland Telegraph, telegraphic officers in South Australia and the Northern Territory were required to report temperatures, barometric pressure, and rainfall on a daily basis to his West Terrace Observatory in Adelaide.

Perhaps as unexpected as the exceptionally long continuous and reliable temperature record for Darwin, Darwin also has the earliest reliable atmospheric pressure measurements. So, the Southern Oscillation Index (SOI) is still measured as the pressure gradient difference—not between Brisbane and Tahiti or Sydney and Tahiti—but between Darwin and Tahiti. These SOI values (expressed as an index) are still derived from the 1887–1989 base period, with the first 10 years of measurements part of the network established by Charles Todd, and still, to this day, updated daily by the BoM.

Changing daily patterns in the SOI were incorporated into the statistical models that John Abbot and I used to forecast monthly rainfall for locations on Australia’s east coast.

There will obviously be problems if rainfall has not been accurately recorded for the location of interest—if the historical record has been corrupted—because the AI will be considering the rainfall total, relative to pressure and temperature gradients and pressures across the Pacific, including at Tahiti.

AI is only as good as the data inputted; AI forecasts are only as good as the data provided for model building, and then for training the model that will be used to make the forecast. Training essentially involves running segments of data to give the model some idea of what to expect. In this regard, AI is like human intelligence: it can get better at anticipating what will happen next, if it is given some practice and good data (reliable information).

ADDITIONAL INFORMATION

*Abbot J., & J. Marohasy, 2012. Application of artificial neural networks to rainfall forecasting in Queensland, Australia. Advances in Atmospheric Sciences, Volume 29, Number 4, Pages 717-730. doi: 10.1007/s00376-012-1259-9

The feature photograph (top of this post) was taken on 22^nd August 2014, for a front-page (if I remember correctly) article in The Australian newspaper by Graham Lloyd detailing our technique using AI for forecasting monthly rainfall up to 18 months in advance.