Multiscale streamflow forecasts for the Brazilian hydropower system using bayesian model averaging (BMA)

ABSTRACT The use of efficient streamflow forecast systems at different time scales allows the operational optimization of the Brazilian interconnected hydropower reservoirs, raising the security level of electricity supply and minimizing operating costs. However, current forecasting models used by the National Electric System Operator (ONS) tend to be limited over the forecast horizon and in the modeling of the dependence structure across the various time scales, thus reducing the quality of forecasts. This paper proposes a new contribution to the streamflow forecast models by exploring the concept of Bayesian Model Averaging (BMA), which allows integrating weekly and monthly forecasts in order to improve the skill of weekly predictions. The monthly forecasts are obtained from a periodic auto-regressive exogenous model (PARX), which attempts to capture the persistence of flow in the auto-regressive part and the runoff contribution in the exogenous portion through the use of climate information. Weekly streamflow forecasts with up to six weeks lead time are obtained from information made available by ONS in the Monthly Operational Program (PMO) reports. The proposed methodology is tested using weekly inflow series from the 28 major Brazilian hydropower reservoirs. The weekly streamflow forecasts results obtained from the weighting of the outputs from the weekly and monthly models indicate a significant improvement in skill based on common performance indicators (NS, MAPE and DM) when compared with forecasts derived from the isolated weekly model. The gains in performance indicators are more significant for lead times beyond two weeks. The proposed approach is flexible in terms of implementation, allowing the incorporation of the other forecast scales as well as different forecast models (e.g. physical models).