Seasonal to interannual rainfall probabilistic forecasts for improved water supply management: Part 1 - A strategy for system predictor identification

Effective use of available water resources is a serious problem facing the world as it enters the 21st century. An important source of concern to water resources managers is the occurrence of severe and sustained droughts that deplete reservoir storage to dangerous levels. Such droughts are often associated with low frequency climatic fluctuations, such as the El Nino Southern Oscillation (ENSO). This paper is part of a study to develop a framework for rainfall probabilistic forecasting using available hydro-climatic information. This paper is the first in a series of three published in this issue, and presents an approach for identifying optimal predictors that can be used to formulate a robust and efficient probabilistic forecast model. The predictor identification approach presented here uses a nonparametric implementation of the mutual information criterion as a measure of dependence between variables. The criterion is based on a characterisation of the joint probability distribution, instead of deviations off a curve of best fit. A partial mutual information criterion is presented as the basis for identifying more than one predictor in a stepwise manner. The method uses nonparametric kernel methods to characterise the joint probability distribution of the variables involved. The method is tested on a range of synthetically generated datasets whose dependence attributes are known beforehand. Results from the application of the partial mutual information criterion to identify predictors of quarterly rainfall using a range of hydro-climatic system variables, are presented in the second paper of this three-paper series. (C) 2000 Elsevier Science B.V. All rights reserved.