Neural network-based metamodelling approach for estimation of air pollutant profiles
- Publication Type:
- Thesis
- Issue Date:
- 2013
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
The air quality system is a system characterised by non-linear, complex
relationships. Among existing air pollutants, the ozone (O3), known as a secondary
pollutant gas, involves the most complex chemical reactions in its formation,
whereby a number of factors can affect its concentration level. To assess the ozone
concentration in a region, a measurement method can be implemented, albeit only at
certain points in the region. Thus, a more complicated task is to define the spatial
distribution of the ozone level across the region, in which the deterministic air
quality model is often used by the authority. Nevertheless, simulation by using a
deterministic model typically needs high computational requirements due to the
nonlinear nature of chemical reactions involved in the model formulation, which is
also subject to uncertainties. In the context of ozone as an air pollutant, the
determination of the background ozone level (BOL), independent from human
activities, is also important as it could represent one of reliable references to human
health risk assessment. The concept of BOL may be easily understood, but
practically, it is hard to distinguish between natural and anthropogenic effects. Apart
from existing approaches to the BOL determination, a new quantisation method is
presented in this work, by evaluating the relationship of ozone versus nitric oxide
(O3-NO) to estimate the BOL value, mainly by using night-time and early morning
measurement data collected at the monitoring stations.
In this thesis, to deal with the challenging problem of air pollutant profile estimation,
a metamodel approach is suggested to adequately approximate intrinsically nonlinear
and complex input-output relationships with significantly less computation.
The intrinsic characteristics of the underlying physics are not assumed to be known,
while the system’s input and output behaviours remain essential. A considerable
number of metamodels approach have been proposed in the literature, e.g. splines,
neural networks, kriging and support vector machine. Here, the radial basis function
neural network (RBFNN) is concerned as it is known to offer good estimation
performance on accuracy, robustness, versatility, sample size, efficiency, and
simplicity as compared to other stochastic approaches. The development
requirements are that the proposed metamodels should be capable of estimating the
ozone profiles and its background level temporally and spatially with reasonably
good accuracies, subject to satisfying some statistical criteria.
Academic contributions of this thesis include in a number of performance
enhancements of the RBFNN algorithms. Generally, three difficulties involved in
the network training, selection of radial basis centres, selection of the basis function
variance (i.e. spread parameter), and training of network weights. The selection of
those parameters is very crucial, as they directly affect the number of hidden
neurons used and also the network overall performance. In this research, some
improvements of the typical RBFNN algorithm (i.e. orthogonal least squares) are
achieved. First, an adaptively-tuned spread parameter and a pruning algorithm to
optimise the network’s size are proposed. Next, a new approach for training the
RBFNN is presented, which involves the forward selection method for selecting the
radial basis centres. Also, a method for training the network output weights is
developed, including some suggestions for estimation of the best possible values of
the network parameters by considering the cross-validation approach. For
applications, results show that the combination of the proposed paradigm could offer
a sub-optimal solution of metamodelling development in the generic sense (by
avoiding the iteration process) for a faster computation, which is essential in air
pollutant profile estimation.
Please use this identifier to cite or link to this item: