Improving monthly rainfall forecasts using artificial neural networks and single-month optimisation: a case study of the Brisbane catchment, Queensland, Australia

Abstract

Official medium-term rainfall forecasts failed to warn of the impending heavy rainfall in the Brisbane catchment during the summer of 2010-2011 with resulting catastrophic flooding causing loss of life, extensive property damage and disruption of economic activity in southeastern Queensland, Australia. Since the flooding, the Australia Bureau of Meteorology has changed its method of forecasting from an empirical statistical scheme to the use of a general circulation model, the Predictive Ocean and Atmospheric Model for Australia (POAMA). More skilful forecasts, however, can be achieved through the use of an altogether different technique involving artificial neural networks (ANN), a form of machine learning. Building on previous studies comparing the skill of the ANN forecasts for Gatton and Harrisville in the Brisbane catchment with output from POAMA, this study shows how monthly ANN rainfall forecasts can be further significantly improved through a technique referred to as "single-month optimisation". This technique enables the temporal variability in the influence of key climate indices to be better incorporated into the rainfall forecast. In particular, correlation coefficients exceeding 0.85 at a 12 month lead time can be achieved for both Harrisville and Gatton.

Forecasting extreme monthly rainfall events in regions of Queensland, Australia using artificial neural networks

“…The present study extends our previous investigations [14], focussing on the capacity of the ANN methodology to differentiate extreme rainfall events from more average conditions. Two regions of Queensland were considered: (i) the coastal region of Queensland extending from Mossman in the far north of the state, through central Queensland, to Maryborough in the south; (ii) the south-eastern region of Queensland, where 54 individual sites were used to generate regional isohyet maps, illustrating the capacity of the methodology to distinguish between extreme and more average rainfall conditions.…”

“…Our previously reported studies [14] focussed on Gatton and Harrisville have reported monthly rainfall forecasts using neural networks. These investigations have demonstrated that using combinations of inputs including both climate indices, and also local variables such as maximum and minimum temperatures [14] produces the best forecasts.…”

“…Based on this formula, the program determines which ANN model and set of inputs is optimal. The present investigation used the same neural network methods previously described for Gatton and Harrisville [14], applied to the individual stations in the two regions under consideration. Each of these selected stations had rainfall record extending back until at least 1920.…”

“…Each of these selected stations had rainfall record extending back until at least 1920. Two approaches have previously been investigated for neural network optimization of monthly rainfall forecasting in Queensland [14]. With the first approach, designated as "allmonth optimization", data for all 12 months of the year were included as input and optimised together, as in our previous studies [8][9][10].…”

“…Our previously reported investigations applied neural networks to individual sites within Queensland with long rainfall records to generate forecasts of monthly rainfall with lead times up to 12 months [8][9][10]. A recent study [14] focussing on the sites of Harrisville and Gatton in south-east Queensland clearly demonstrated that superior monthly rainfall forecasts can be generated using the ANN approach when compared to results generated by POAMA. Forecasts generated using the ANN approach were consistently more skilful than climatology, as shown by lower root means square error (RMSE) and mean absolute error (MAE) values.…”