Probabilistic Estimation of Multivariate Streamflow Using Independent Component Analysis and Climate Information

Abstract

A statistical estimation approach is presented and applied to multiple reservoir inflow series that form part of Sydney?s water supply system. The approach involves first identifying sources of interannual and interdecadal climate variability using a combination of correlation- and wavelet-based methods, then using this information to construct probabilistic, multivariate seasonal estimates using a method based on independent component analysis (ICA). The attraction of the ICA-based approach is that, by transforming the multivariate dataset into a set of independent time series, it is possible to maintain the parsimony of univariate statistical methods while ensuring that both the spatial and temporal dependencies are accurately captured. Based on a correlation analysis of the reservoir inflows with the original sea surface temperature anomaly data, the principal sources of variability in Sydney?s reservoir inflows appears to be a combination of the El Niño?Southern Oscillation (ENSO) phenomenon and the Pacific decadal oscillation (PDO). A multivariate ICA-based estimation model was then used to capture this variability, and it was shown that this approach performed well in maintaining the temporal dependence while also accurately maintaining the spatial dependencies that exist in the 11-dimensional historical reservoir inflow dataset.