City University of Hong Kong
DSpace
 

CityU Institutional Repository >
3_CityU Electronic Theses and Dissertations >
ETD - Dept. of Building and Construction  >
BC - Doctor of Philosophy  >

Please use this identifier to cite or link to this item: http://hdl.handle.net/2031/4922

Title: Multilayer perceptron models for surface ozone study in Hong Kong under the trans-boundary air pollution impact
Other Titles: Zai kua jie kong qi wu ran ying xiang xia yong gan zhi qi mo xing dui Xianggang xiu yang zhi yan jiu
在跨界空氣汚染影響下用感知器模型對香港臭氧之研究
Authors: Wang, Dong (王東)
Department: Dept. of Building and Construction
Degree: Doctor of Philosophy
Issue Date: 2007
Publisher: City University of Hong Kong
Subjects: Atmospheric ozone -- China -- Hong Kong
Perceptrons
Notes: CityU Call Number: Q327.W36 2007
Includes bibliographical references (leaves 217-227)
Thesis (Ph.D.)--City University of Hong Kong, 2007
xxi, 240 leaves : ill. ; 30 cm.
Type: Thesis
Abstract: Multilayer perceptron (MLP) models have been experiencing a popularity resurgence for predicting surface ozone level, based on the data of its influential variables colleted locally from air quality monitoring and meteorology stations around the interested area. In this dissertation, MLP, not only used traditionally as a predictive model but also an assessment tool, will be employed to study ozone variation in three typical air monitoring stations in Hong Kong, the ozone variation of where are believed to be under different scale of trans-boundary air pollution impact. The optimal topology of each MLP model used for assessment or prediction was identified by 3-fold cross validation for two prediction horizons respectively. For assessment work, result from both prediction horizons shows no remarkable difference. While for prediction work, performance of all MLP models for 1-day ahead prediction was generally worse than that for the current-day prediction due to the prolonged prediction horizon. The preliminary statistical analysis showed the trans-boundary air pollution did exert different scale of influence on local ozone level in each target air monitoring station, according to the data for all local and regional ozone influential variables collected from the whole study area that are defined as Hong Kong territory, Guangdong Province and part of South China Sea. MLP trained by automatic relevance determination (named by MLP-ARD), a Bayesian MLP, was embedded into a two-staged variables selection scheme to assess what were the ozone key influential variables for each air monitoring station respectively. The variables selection/assessment result from MLP-ARD was comparable with that from the best method in the literature. The ozone key influential variables from such selection scheme will further be used as input variables for MLP models developed later for ozone prediction. By comparing the predictive performance of MLP-ARD between with and without regional ozone influential variables as inputs, it was found that trans-boundary air pollution exerted the largest impact on Tap Mun (TM), the modest on Tung Chung (TC), and the least on Tsuen Wan (TW) air monitoring station in Hong Kong. The result also showed the advantage of MLP-ARD, which provided an interval estimation for the possible ozone variation, in the prediction for ozone episode days, over the MLP trained by Levenberg-Marquardt (LM) algorithm (named by MLP-LM), which only provided a point estimation. MLP-ARD, MLP-LM as well as most MLP models in the literature were trained by the gradient-based algorithm, which potentially suffered from local minimum problem. Two hybrid MLP models, based on the standard particle swarm optimization (PSO) algorithm, will be developed for avoiding this problem in ozone prediction. The reason for using hybrid model instead of using MLP trained by standard PSO (named by MLP-PSO) directly is that standard PSO for MLP training will probably not obtain good convergence reliability in the high dimension weight space, which could influence the models‘ performance on the ozone-polluted days. Therefore, the aim of two hybrid models was to improve such convergence reliability by using additional techniques before and after the standard PSO training for MLP respectively. The first hybrid model was HMC–MLP–PSO. It employed hybrid Monte Carlo (HMC) method to sample the weight matrix from the posterior probability distribution of the estimated optimal weight matrix first, and then these sampled weight matrices were used to initialize ―weight matrix swarm‖ of PSO, before MLP trained by standard PSO starts. Aiming at exploiting the advantage of both PSO and LM for MLP training, the other hybrid model was MLP–PSO–LM, which timely inserted LM to help MLP-PSO avoid stagnation problem. The performance of two hybrid models was better than MLP-ARD, MLP-LM and MLP-PSO in terms of several statistics and exceedance indicators. The predictive performance of all MLP models in this dissertation was finally evaluated. From the operational point of view, MLP–PSO–LM was recommended for government authority usage due to its smallest false negative rate for ozone-polluted day prediction.
Online Catalog Link: http://lib.cityu.edu.hk/record=b2218227
Appears in Collections:BC - Doctor of Philosophy

Files in This Item:

File Description SizeFormat
fulltext.html157 BHTMLView/Open
abstract.html157 BHTMLView/Open

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

 

Valid XHTML 1.0!
DSpace Software © 2013 CityU Library - Send feedback to Library Systems
Privacy Policy · Copyright · Disclaimer