Climate informed monthly streamflow forecasts for the Brazilian hydropower network using a periodic ridge regression model

Streamflow simulation and forecasts have been widely used in water resources management, particularly for flood and drought analysis and for the determination of optimal operational rules for reservoir systems used for water supply and energy production. Here we include climate information in a periodic-auto-regressive model in order to provide monthly streamflow forecasts for 54 hydropower sites in Brazil. Large scale climate information is included in the model through the use of climate indices obtained from the sea surface temperature field of the tropical Pacific and sub-tropical Atlantic oceans and the low-level zonal wind field over southeast Brazil. Correlation analysis of climate predictors and streamflow data show that the dependence of the latter on climate variability is seasonal and also a function of the lead time of the forecasts. A ridge regression framework is adopted in order to shrink parameter estimates and improve model outputs. The proposed model is compared with an ordinary linear regression based model with predictors selected by the BIC criterion and with the classical linear periodic-auto-regressive model (PAR), where no climate information is used. Cross-validated results show that the inclusion of climate indexes is able to improve forecast skills up to 3 months lead time. Higher skills are observed for reservoirs with large catchment areas.