HR: 1340h
AN: H13G-1396    [Abstracts]
TI: Using Independent Component Analysis for Medium-Term Forecasts of Rainfall at Multiple Sites.
AU: * Westra, S
EM: seth@civeng.unsw.edu.au
AF: School of Civil and Environmental Engineering, University of New South Wales , Sydney, NSW 2052 Australia
AB: The development of an appropriate regression model to forecast regional hydro-climatic variables such as rainfall and stream-flow for monthly through to annual time scales has been the subject of significant research in recent decades. This research has focused largely on trying to overcome four important challenges related to the complex nature of the variables involved; namely, (1) identifying a parsimonious set of predictors out of a range of potential predictor datasets such as sea surface temperatures (SST) or sea level pressures (SLP), which are highly correlated in space and time and are of very high dimension, (2) capturing the statistical dependence between the predictor and predictand, which may be linear or non-linear, (3) capturing the spatial dependence within the predictand dataset, where a high level of spatial correlation is often present, and (4) characterising the uncertainty of the ensuing forecast. To address these challenges, a forecasting method is described which uses the principles of a recently developed component extracting technique known as independent component analysis (ICA). This method may be regarded as an extension to canonical correlation analysis (CCA) which has been used extensively to forecast variables such as rainfall at multiple sites, except that while CCA considers only second-order dependence, the method proposed here also considers higher-order statistical information. Thus, whereas CCA seeks to maximise the correlation between linear combinations of the predictor and predictand datasets, the ICA-based method seeks to maximise the mutual information between these datasets, thereby also taking into account any non-linear relationships. Furthermore, whereas successive canonical variables are constrained to be mutually uncorrelated, the ICA approach focuses on statistical independence, which is a much more stringent mathematical criterion. The ICA approach not only has significant potential in improving the forecasts that are obtained, but is also likely to lead to results that are physically interpretable. The advantages of the ICA approach will be illustrated using a synthetic example, and the influence of several parameters governing the implementation of the approach will be discussed. This will then be followed by the application of the method to forecasting regional rainfall, and conclusions will be drawn on the predictive capacity and interpretability of the results.
DE: 1816 Estimation and forecasting
DE: 1833 Hydroclimatology
DE: 1854 Precipitation (3354)
DE: 1869 Stochastic hydrology
DE: 4215 Climate and interannual variability (1616, 1635, 3305, 3309, 4513)
SC: Hydrology [H]
MN: Fall Meeting 2005