Please use this identifier to cite or link to this item: http://hdl.handle.net/10397/77368
Title: Hybrid model of neural network and population-based optimization algorithm for river flow and sediment load
Authors: Chen, Xiaoyun
Advisors: Chau, K. W. (CEE)
Keywords: Streamflow
Sediment transport
River sediments
Issue Date: 2018
Publisher: The Hong Kong Polytechnic University
Abstract: River flow forecasting and sediment load estimation are important issues in the hydrological field, and customarily undertaken by data-driven models. Despite the amount of research on the subject, most of them are incapable of providing insight into the unrecognized relationship of the input and output variables owing to a black-box nature. It is inevitable that the physically-meaningless models suffer from inappropriate input-output mapping, substantial uncertainty inherent in the modeling and inadequate optimization during calibration. This thesis is an attempt to develop physics-based models for efficiently and reliably simulating the river flow and sediment load. A novel hybrid neural network (HNN) model is proposed for downstream river flow forecasting, by combining fuzzy pattern-recognition and continuity equation in a neural network. The model is therefore, able to reflect fuzzy and time-varying features of river flows. In comparison with three benchmarking models, the HNN model is identified as the preferred tool to fit the total observations. The superiority of HNN model does not markedly deteriorate with the increase of forecasting lead time. With respect to sediment load estimation, a hybrid double feedforward neural network (HDFNN) model is developed by integrating fuzzy pattern-recognition and continuity equation into a structure of double neural networks. The generalization and estimation abilities of HDFNN models are verified by comparison with results from its counterpart models. It could reproduce medium and high loads appropriately, and present excellent performances in multi-step-ahead estimations particularly for high loads. The request of determining the best input variables for the HNN and HDFNN models has been processed. Generally, areal precipitation is an appropriate input variable coupled with all observed upstream flows for the HNN model. The input variables may be uncertain and unstable when forecast lead time is shorter than the flow travel time. For the sediment estimation, river flows from the upstream and downstream stations with different ahead of times are selected to formulate input combinations. It is found that the downstream sediment seems to be more sensitive to upstream flows with small studied area while the downstream flows substantially affects the high sediment loads. The lower upper bound estimation (LUBE) is a straight-forward method that could construct the neural network based models with two output neurons and directly approximate the lower and upper bounds of prediction intervals. Applications on the HNN and HDFNN models indicate their reliability in hydrological prediction scenarios. The performances of three population-based optimization algorithms, namely differential evolution (DE), artificial bee colony (ABC) and ant colony optimization (ACO) for evolving the HNN and HDFNN models are compared. The DE is found to be a more appropriate algorithm in terms of generalization and prediction. The ABC appears to be more adapted in optimizing the multi-step-ahead cases, but on the other hand, presents computational inefficiency. As far as the stability is concerned, the DE and ABC algorithms are more adaptive than the ACO with the population size. The major contribution of this research is the development of HNN and HDFNN models for river flow and sediment load. The LUBE method has proven as a promising technique to evaluate the model reliability. In addition, this thesis advocates the use of DE algorithm for the optimization problems in hydrological models.
Description: xxi, 269 pages : color illustrations
PolyU Library Call No.: [THS] LG51 .H577P CEE 2018 Chen
URI: http://hdl.handle.net/10397/77368
Rights: All rights reserved.
Appears in Collections:Thesis

Files in This Item:
File Description SizeFormat 
991022131142603411_link.htmFor PolyU Users167 BHTMLView/Open
991022131142603411_pira.pdfFor All Users (Non-printable)3.41 MBAdobe PDFView/Open
Show full item record
PIRA download icon_1.1View/Download Contents

Page view(s)

19
Citations as of Dec 9, 2018

Download(s)

5
Citations as of Dec 9, 2018

Google ScholarTM

Check


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