Please use this identifier to cite or link to this item: http://hdl.handle.net/10397/40928
Title: Improving neural network river forecasting with swarm-based optimization algorithms
Authors: Taormina, Riccardo
Advisors: Chau, K. W. (CEE)
Keywords: Stream measurements.
Neural networks (Computer science)
Mathematical optimization.
Issue Date: 2016
Publisher: The Hong Kong Polytechnic University
Abstract: Neural Network River Forecasting (NNRF) entails the use of Artificial Neural Networks (ANNs) for the prediction of streamflow quantities. Despite the amount of research on the subject, NNRF still struggle to move from the academic context to the operational context due to a number of unresolved issues. Major problems of NNRF methodologies include a) difficulties in quantifying the uncertainty of model predictions, b) the lack of standardized methodologies for identifying optimal predictors and suitable functional forms of the underlying data-driven model, and c) concerns with the black-box nature of NNRF models which drive practitioners to favour physically-based alternatives. The main contribution of this thesis is to show that these issues, albeit very different in nature, can all be addressed by developing NNRF models using Global Optimization. In particular, this work introduces three new Particle Swarm Optimization (PSO) variants which are employed to devise novel ad-hoc applications aimed at solving each particular issue. These algorithms are the Multi-Objective Fully Informed Particle Swarm (MOFIPS) optimization, the Binary-coded Fully Informed Particle Swarm (BFIPS), and its multi-objective generalization (MBFIPS). Testing these new techniques will also provide insights on the real effectiveness of PSO for data-driven hydrological modelling, a task which has been only partially accomplished by the research community. In addition, this thesis advocates the use of Extreme Learning Machines (ELMs) as alternative NNRF models. Although research in other fields has shown that ELMs provides better accuracy at much faster speed compared to ANNs, at the time of writing, they have never been employed for NNRF modeling.
There are four applications at the core of this thesis. In a first application it is demonstrated that better deterministic PSO-trained NNRF models can be obtained by formulating cross validation as a bi-objective optimization problem using MOFIPS to perform ANN calibration. The benefits of bi-objective optimization are also shown for the construction of NNRF prediction intervals. This is done in a second application where MOFIPS and the Lower Upper Bound Estimation method are employed for fast and straightforward development of interval-based models. In a third study, a novel approach for model and Input Variable Selection (IVS) that employs BFIPS and MBIFPS along with the ELMs is presented. A comparison with 4 existing techniques, done using the tools of a comprehensive framework, suggests that the developed ELM-based models are more accurate in performing the IVS task for data-driven hydrological modelling. Lastly, BFIPS, MBFIPS and ELMs are employed to investigate whether more accurate prediction of streamflow discharges can be achieved by including expert knowledge in NNRF model development. In particular, total streamflow predictive accuracy of modular models (MM) trained to perform an implicit baseflow separation is compared against that of global models (GM). The results for 9 different watersheds in northern United States show that MMs underperform GMs in predicting the total flow. In addition, the study demonstrates that greater accuracy in baseflow separation usually corresponds to worse total flow predictions, suggesting that these two objectives are conflicting, rather than compatible.
Description: PolyU Library Call No.: [THS] LG51 .H577P CEE 2016 Taormina
xi, 169 pages :illustrations
URI: http://hdl.handle.net/10397/40928
Rights: All rights reserved.
Appears in Collections:Thesis

Files in This Item:
File Description SizeFormat 
b28563864_link.htmFor PolyU Users208 BHTMLView/Open
b28563864_ira.pdfFor All Users (Non-printable)5.68 MBAdobe PDFView/Open
Show full item record

Page view(s)

72
Last Week
2
Last month
Checked on Jun 25, 2017

Download(s)

43
Checked on Jun 25, 2017

Google ScholarTM

Check



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