Details

Autor(en) / Beteiligte

• Panahi, Javad
• Mastouri, Reza
• Shabanlou, Saeid

Titel

Insights into enhanced machine learning techniques for surface water quantity and quality prediction based on data pre-processing algorithms

Ist Teil von

• Journal of hydroinformatics, 2022-07, Vol.24 (4), p.875-897

Ort / Verlag

London: IWA Publishing

Erscheinungsjahr

2022

Link zum Volltext

Quelle

Electronic Journals Library - Freely accessible e-journals

Beschreibungen/Notizen

• Abstract Quality and quantity of streamflow are crucial components in the management and control of water resources according which are challenging due to their nonstationarity and uncertainty path. This paper presented an ensemble data pre-processing-based machine learning (ML) algorithm for the decision-support of water resource management and water pollution control at the watershed scale due to the nonlinear path of streamflow. In the proposed hybrid model, a new time–frequency analysis algorithm, variational mode decomposition (VMD), is implemented to deal with the nonlinearity and nonstationary of a streamflow process. The VMD is exploited to decompose the original water quality and quantity series into a series of intrinsic mode functions (IMFs) with different frequencies. Therefore, an ensemble algorithm, bootstrap aggregating (bagging) algorithm is coupled with two common ML, reduced error pruning tree (REPT) and random forest (RF), to predict all the decomposed modes using VMD. Then, in order to reduce the variance among the base classifiers of the proposed ML, a bootstrap aggregation technique was recruited. Finally, the predicting value of the original water quality and quantity series is obtained by adding up the predicting results of all the decomposed modes. The proposed hybrid decomposition–ensemble model has been applied to two stations in Karoon River, Iran. Results obtained from this study indicate that the proposed hybrid decomposition–ensemble model can capture the nonlinear characteristics of a streamflow process in terms of water quality and quantity simultaneously and thus provide more accurate predicting results compared with those models without data frequency decomposing.