Weekly Streamflow Forecasting Using a Statistical Disaggregation Model for the Upper Blue Nile Basin, Ethiopia

Accurately forecasting the weekly seasonal streamflow of the Upper Blue Nile basin (UBNB) in Ethiopia is essential for managing large-scale water projects of Nile basinwide countries. A wavelet-based, artificial neural network calibrated by genetic algorithm (ANN-GA) model and a statistical disaggregation algorithm were integrated to forecast weekly streamflow of the UBNB. The July to October (JASO) streamflow of the El Diem station of UBNB shows strong interannual oscillations prior to the 1920s and after 1990s. Two ANN-GA models were developed to forecast the UBNB JASO streamflow, the first one using the February to May (FMAM) seasonal sea surface temperature (SST) of the global oceans as predictors to directly forecast JASO streamflow, while the second, a hybrid model, is developed to forecast JASO streamflow from two sets of predictors, which consist of FMAM SST and the July to September (JJAS) seasonal rainfall previously forecasted by the wavelet-based, ANN-GA also driven by FMAM SST as predictors. The forecasted JASO streamflow were then disaggregated to weekly total streamflow using the disaggregation model, Valencia and Schaake (VS). Results indicate that seasonal forecasts with up to 4months lead time only based on SST as predictors achieved reasonable skill (r2=0.5), while the hybrid model achieved a better performance (r2=0.8). The disaggregated streamflow from the first model explained 69% of the streamflow variance, compared to 71% when the hybrid model was used. Based on the results, the proposed hybrid model that uses both SST and a forecasted JJAS seasonal rainfall as predictors achieves a marginally better forecasting skill of the UBNB weekly streamflow. This proposed method that directly forecasts the streamflow, rather than forcing a hydrological model with rainfall forecasts is useful for the management of river basins that lack observed hydrologic data.