A Hybrid DBN Model Framework for Long-term Streamflow Forecasts

A reliable long-term streamflow forecast is crucial for water resources management and reservoir operations.However,the accuracy of long-term streamflow forecasts are usually unsatisfactory and incapable of meeting stakeholders’ needs.In this study,a long-term streamflow forecasting framework based on the deep belief network (DBN) model was proposed and applied to the Tianyi reservoir for monthly streamflow forecasts up to 12 months ahead.The results show that the proposed framework produces reasonable performance with Nash-Sutcliffe efficiency coefficient (NSE) larger than 0.50 and mean relative error (MRE) lower than 35% for all lead months.In addition,the forecasting performance keeps in stable with the extension of the lead time.Moreover,the framework exhibits good generalization ability with no obvious difference between the training and testing periods.Finally,when compared with non-flood season,the forecasting framework exhibits higher NSE and R² values as well as larger MRE values in the flood season for all lead times.