A physics-informed statistical learning framework for forecasting local suspended sediment concentrations in marine environment. (30th June 2022)

• Record Type: Journal Article
• Authors: Zhang, Shaotong; Wu, Jinran; Wang, You-Gan; Jeng, Dong-Sheng; Li, Guangxue
• Is Part Of: Water research. Volume 218(2022)
• Publication Date: 2022-06-30
• Subjects: Water quality -- Marine ranching -- Temporal auto-correlation -- Augmented lncosh ridge regression -- Outlier handling -- The Yellow River Delta; Water -- Pollution -- Research -- Periodicals; 363.7394
• DOI: 10.1016/j.watres.2022.118518
• Languages: English
• ISSNs: 0043-1354

Abstract

Highlights: A decomposition-ensemble framework is built for single-parameter modeling of superposed periodic signals. An augmented lncosh ridge regression is proposed to handle the outliers and consider the static settling of suspended sediments. A case study shows the framework realized 6-hour ahead SSC forecast with MREs of 5.80–9.44%. Graphical abstract: Abstract: An in-situ monitoring of water quality (suspended sediment concentration, SSC) and concurrent hydrodynamics was conducted in the subaqueous Yellow River Delta in China. Empirical mode decomposition and spectral analysis on the SSC time series reveal the different periodicities of each physical mechanism that contribute to the SSC variations. Based on this physical understanding, the decomposed SSC time series were trained separately with a newly-proposed augmented lncosh ridge regression, in which (1) a lncosh function was incorporated in traditional ridge regression for handling outliers in original data, and (2) the temporal auto-correlation in the decomposed SSC series was used for augmented regression. Finally, the trained sub-series were added up as the final prediction. The advantages of this decomposition-ensemble framework is that it depends on SSC only, superior to the normal process-based models which need the concurrent hydrodynamics for estimating bed shear stress. This will not only reduce the measurement uncertainties of the input when training the data-driven model, but also save the prediction cost as no other parameters than SSC need to be measured and input for running the model. The framework realized 6-hour-ahead high-accuracy forecasting with mean relative errors of 5.80–9.44% in the present case study. The proposed framework can be extended to forecast any signal that is superposed by components with various timescales (periodicities) which is common in nature.