Monthly Streamflow Forecasting Using EEMD-Lasso-DBN Method Based on Multi-Scale Predictors Selection

Abstract

For the inherent characteristics of a raw streamflow times series and the complicated relationship between multi-scale predictors and streamflow, monthly streamflow forecasting is very difficult. In this paper, an method was proposed integrating the ensemble empirical mode decomposition (EEMD), least absolute shrinkage and selection operator (Lasso) with deep belief networks (DBN) for forecasting monthly streamflow time series, which is EEMD-Lasso-DBN (ELD) method. To develop the ELD model, the raw streamflow time series was resolved into different elements, including intrinsic mode functions (IMFs) and residue series, using the EEMD technique. The predictors of each IMF element and residue were screened using the Lasso technique from a large number of candidate predictors, respectively. Then, the DBN models were built to simulate the complex relationship between the resolved elements and the selected predictors, respectively. The predicted results of the IMFs and residual series were assembled as an ensemble forecast for the raw streamflow time series and were compared with the other models. The monthly streamflow series from Tennessee, in the USA, were investigated using the ELD method. It was found that each IMF has different characteristics and physical meaning, corresponding to different predictors. The proposed ELD model can significantly improve the accuracy of monthly streamflow forecasting.