103-Coping with Model Structural Uncertainty in Medium Term Probabilistic Streamflow Forecasting Applications.

Abstract

This paper describes the rationale behind the Multiple Reservoir Inflow Forecasting System and how it is specified to issue 12 months long probabilistic forecasts across selected reservoir systems in Australia. The forecasting system enables development of 12 months long monthly flow scenarios (probabilistic forecasts) issued 'based on a mix of prevalent catchment and climate conditions, so as to simulate the likely evolution of the reservoir network over the coming 12 months for medium term planning and operations. The probabilistic forecasting system uses previous season flow as an indicator of catchment influences and selected sea surface temperature anomalies across varying locations and time lags as climatic indicators in its formulation. Given the considerable uncertainty that exists in choosing relevant climatic indicators, a flexible model structure (or multiple climatic indicators) is adopted, leading to the formulation of multiple predictive models, each aiming to represent the impact different climate phenomena have on the responses. Examples of such multiple models could be an El Nino Southern Oscillation indicator at a short time lag for one model, and a combination of the Inter-decadal Pacific Oscillation and a lagged Indian Ocean dipole influence for another model. The inflow scenarios that are finally predicted thus come from a collection of such models, thereby enabling an appropriate representation of the structural uncertainty that exists in this prediction problem. The above forecasting system is illustrated using two case studies representing the Sydney Catchment Authority water supply system and the Tasmanian hydroelectric system. The advantages and drawbacks of using the above mentioned model averaging logic are discussed, and modifications suggested that can improve the predictive performance and consequent reservoir operations.