Using model averaged probabilistic forecasts for water resources management

A drawback of medium-to-long-term probabilistic forecasting methods is the relatively high uncertainty associated with model outputs, particularly when. the models are used for prediction of future scenarios. This paper presents an extension to the probabilistic forecasting approach first presented in Sharma (2000b), that attempts to enhance the reliability of the model using an ensemble averaging approach. Each ensemble member or model is formulated using nonparametric statistical techniques and is restricted to have a relatively independent basis so as to represent the multiple mechanisms that influence the system being studied. The aim of using ensemble or model averaging is to reduce the chance of model misspecification, a common occurrence when the dependence is highly random and the system too complex to be explained by a limited number of predictors. The usefulness of the procedure is demonstrated through an application to forecast the Southern Oscillation Index (SOI), the multiple models being formulated using predictors selected from prior lags of the SOI and globally distributed, gridded sea surface temperature anomaly data. The model is assessed by evaluating its performance both in cross-validation as well as by forecasting an entire period of the record that was left out in the model formulation process. The results indicate that the consideration of uncertainty in climatological observations and the use of an ensemble of model outputs results in probabilistic forecasts that are more reliable and accurate than is the case otherwise. The implications of using the probabilistic forecasts for water resources management are discussed.