Please use this identifier to cite or link to this item: http://hdl.handle.net/10397/43768
Title: Data-driven input variable selection for rainfall-runoff modeling using binary-coded particle swarm optimization and Extreme Learning Machines
Authors: Taormina, R
Chau, KW 
Keywords: Extreme learning machines
Input variable selection
Multi-objective optimization
Particle swarm optimization
Rainfall-runoff
Streamflow prediction
Issue Date: 2015
Publisher: Elsevier
Source: Journal of hydrology, 2015, v. 529, p. 1617-1632 How to cite?
Journal: Journal of hydrology 
Abstract: Selecting an adequate set of inputs is a critical step for successful data-driven streamflow prediction. In this study, we present a novel approach for Input Variable Selection (IVS) that employs Binary-coded discrete Fully Informed Particle Swarm optimization (BFIPS) and Extreme Learning Machines (ELM) to develop fast and accurate IVS algorithms. A scheme is employed to encode the subset of selected inputs and ELM specifications into the binary particles, which are evolved using single objective and multi-objective BFIPS optimization (MBFIPS). The performances of these ELM-based methods are assessed using the evaluation criteria and the datasets included in the comprehensive IVS evaluation framework proposed by Galelli et al. (2014). From a comparison with 4 major IVS techniques used in their original study it emerges that the proposed methods compare very well in terms of selection accuracy. The best performers were found to be (1) a MBFIPS-ELM algorithm based on the concurrent minimization of an error function and the number of selected inputs, and (2) a BFIPS-ELM algorithm based on the minimization of a variant of the Akaike Information Criterion (AIC). The first technique is arguably the most accurate overall, and is able to reach an almost perfect specification of the optimal input subset for a partially synthetic rainfall-runoff experiment devised for the Kentucky River basin. In addition, MBFIPS-ELM allows for the determination of the relative importance of the selected inputs. On the other hand, the BFIPS-ELM is found to consistently reach high accuracy scores while being considerably faster. By extrapolating the results obtained on the IVS test-bed, it can be concluded that the proposed techniques are particularly suited for rainfall-runoff modeling applications characterized by high nonlinearity in the catchment dynamics.
URI: http://hdl.handle.net/10397/43768
ISSN: 0022-1694
DOI: 10.1016/j.jhydrol.2015.08.022
Appears in Collections:Journal/Magazine Article

Access
View full-text via PolyU eLinks SFX Query
Show full item record

SCOPUSTM   
Citations

66
Last Week
1
Last month
Citations as of Sep 10, 2017

WEB OF SCIENCETM
Citations

61
Last Week
0
Last month
Citations as of Sep 15, 2017

Page view(s)

33
Last Week
1
Last month
Checked on Sep 18, 2017

Google ScholarTM

Check

Altmetric



Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.