River Suspended Sediment Prediction Using Various Multilayer Perceptron Neural Network Training Algorithms—A Case Study in Malaysia

• Published in: Water resources management Vol. 26; no. 7; pp. 1879 - 1897
• Main Authors: Mustafa, M. R., Rezaur, R. B., Saiedi, S., Isa, M. H.
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01.05.2012; Springer; Springer Nature B.V
• Subjects: Algorithms; Atmospheric Sciences; Case studies; Civil Engineering; Convergence; Discharge; Earth and Environmental Science; Earth Sciences; Earth, ocean, space; Environment; Exact sciences and technology; Freshwater; Geotechnical Engineering & Applied Earth Sciences; Hydraulic structures; Hydrogeology; Hydrology. Hydrogeology; Hydrology/Water Resources; Load; Malaysia; Neural networks; Neurons; Performance evaluation; prediction; problem solving; Rivers; Sediment discharge; Sediment transport; sediment yield; Sediments; Stream flow; Suspended sediments; Training; Water resources; Far East; Asia; Malaysia; Suspended sediment; Multilayer perceptron neural network; Training algorithms; Modeling; Discharge; Prediction; rivers; algorithms; neural networks; accuracy; case studies; suspended materials; hydraulics; water resources; Discharge · Suspended sediment; discharge; water resource management; prediction; s Multilayer perceptron neural network
• ISSN: 0920-4741; 1573-1650
• DOI: 10.1007/s11269-012-9992-5

Estimation of suspended sediment discharge in rivers has a vital role in dealing with water resources problems and hydraulic structures. In this study, a Multilayer Perceptron (MLP) feed forward neural network with four different training algorithms was used to predict the suspended sediment discharge of a river (Pari River at Silibin) in Peninsular Malaysia. The training algorithms are Gradient Descent (GD), Gradient Descent with Momentum (GDM), Scaled Conjugate Gradient (SCG), and Levenberg Marquardt (LM). Different statistical measures, time of convergence and number of epochs to reach the required accuracy were used to evaluate the performance of training algorithms. The analysis showed that SCG and LM performed better than GD and GDM. While the performance of the superior algorithms (i.e., SCG and LM) is similar, LM required considerably shorter time of convergence. It was concluded that both training algorithms SCG and LM could be recommended for suspended sediment prediction using MLP networks. However, LM was the faster (1/7 of SCG convergence time) of the two algorithms.