Application of the novel circulant singular spectrum analysis ensemble model for forecasting of streamflow data

Abstract

Accurate forecasting of streamflow is an important part of water resource management and a difficult job for engineers. In this study, a circulant singular spectrum analysis (CiSSA)-based novel approach for forecasting daily streamflow data is proposed. Obtained features using CiSSA methods are applied to support vector regression (SVR), random forest (RF), and artificial neural network (ANN) models. In addition, the variational mode decomposition (VMD) method was used for preprocessing the streamflow data and to compare the performance of the CiSSA method.

One to five-day ahead forecasting performance was investigated and compared with single SVR, RF, ANN single models. VMD-SVR, VMD-RF, VMD-ANN, CiSSA-SVR, CiSSA-RF, CiSSA-ANN, CiSSA-VMD-SVR, CiSSA-VMD-RF, and CiSSA-VMD-ANN ensemble models and Kruskal–Wallis test was used to show whether the forecasting results are statically significant. The analysis of the obtained results showed that the performance of the models (CiSSA-SVR, CiSSA-RF, and CiSSA-ANN) that use CiSSA features as input for far-day ahead predictions were significantly better than the other models. Taylor diagrams and mean absolute error (MAE), root mean square error (RMSE) and correlation coefficient (R), Nash–Sutcliffe efficiency coefficient (NSE), and Willmott’s refined index (WI) performance parameters revealed that the performance of the CiSSA-RF model was better. Furthermore, the CiSSA-RF model proved to be significantly much reliable for longer predicting periods.