Modelling the Variability of Rainfall Extremes - a Case Study for Sydney, Australia

Evidence is mounting that the assumption of non-stationarity in modelling rainfall extremes is no longer valid but one has to account for climate change and climate variability. In this case study, the emphasis is on climate variability because the variability in rainfall extremes exceeds the magnitude of the observed long-term trend. The statistical modelling of rainfall extremes is based on an artificial neural network approach developed by Cannon (2010) that allows accounting for non-linear interactions between covariates. The variability of rainfall extremes needs to be discussed in the context of temporal and spatial scales. On the global scale, phenomena like the El Nino-Southern Oscillation (ENSO) and the Interdecadal Pacific Oscillation (IPO) govern rainfall variability over months and years, while variability on shorter timescales is related to synoptic and mesoscale phenomena. We discuss the selection of suitable covariates from a set of large-scale drivers (El Nino-Southern Oscillation, Indian Ocean Dipole, Southern Annular Mode, Interdecadal Pacific Oscillation) and provide context by discussing other factors affecting the occurrence of rainfall extremes in the Sydney region (synoptic situations and in particular East Coast Lows). We compare our results to those from a stationary Generalised Extreme Value model. Our analyses indicate that in this case one has to include more than one covariate to model the non-stationarity due to climate variability. It is found that due to interactions between covariates the effects on rainfall extremes are not linear and require sophisticated modelling techniques. Comparison of our results with those from the stationary model indicate that confidence intervals derived based on the assumption of stationarity may be too narrow and a better approximation of uncertainty requires a more complex model. This approach could be extended to other locations where the variability in rainfall extremes is significant when compared to the long-term trend. Where operational seasonal forecasts have sufficient skill in predicting large-scale drivers (e.g. ENSO) for the season ahead, this approach may have relevance for practical applications.