Keywords: air pollutant, artificial neural network, forecast, radial basis function, support vector machine,.

Forecasting of air quality parameters is an important topic of atmospheric and environmental research today due to the health impact caused by airborne pollutants existing in urban areas. As the continuous economy development and the population increase in metropolitan city like Hong Kong, Shanghai, etc., a series of severe problems relating to the environmental protection have attracted much attention than ever before, e.g., air pollution, noise pollution, shortage of land resources, waste and sewage disposal, etc. Among these, air pollution has direct impact on human health through exposure to pollutants at high concentration level existing in ambient. Air pollution control is needed to prevent the situation getting worse in the long run. On the other hand, short-term forecasting of air quality is needed in order to take preventive and evasive action during episodes of airborne pollution.

Environmental modelling involves using a variety of approaches. The use of neural networks, and, in particular, the multi-layer feed-forward neural networks, which can be trained to approximate virtually any smooth, measurable function, have become popular and produced promising results. A detailed review of the application of multi-layer feed-forward neural networks in atmospheric science is also available [1]. This research has shown that the neural network approach is effective in simulating and describing the dynamics of non-stationary time series due to its unique non-parametric, non-assumable, noise-tolerant and adaptive properties.

However, a critical issue concerning the neural networks is the over-fitting problem. It can be attributed to the fact that a neural network, sometimes, captures not only useful information contained in the given data, but also unwanted noise. This usually leads to a poor level of generalization. The performance of neural networks in terms of generalization for the out-of-sample data - the data are not used in training the network - is always inferior to that of the training data. Therefore, the development of neural networks and the tasks related to architecture selection, learning parameters estimation and network training, require excessive care in order to achieve the desired level of generalization.

The support vector machine (SVM), developed by Vapnik [2], have provided a new, effcetive novel approach to improve the generalization of neural networks. Originally, SVMs were developed for pattern recognition problems. Recently, with the introduction of -insensitive loss function, SVMs have been extended to solve non-linear regression estimation and time series prediction [3,4]. Unlike most of the traditional learning machines, which adopt the Empirical Risk Minimization Principle (ERMP), e.g., feed-forward neural network, SVMs implement the Structural Risk Minimization Principle (SRMP), which seeks to minimize an upper bound of the generalization error rather than minimize the training error. Such process results in better generalization than conventional techniques.

This paper presents a pioneer study of using SVM to forcast the concentration variations of six air pollutants hourly measured during the whole year of 1999 at the Causeway Bay Roadside Gaseous Monitory Station, one of the fourteen pollutant monitory stations established by Hong Kong Environment Protection Department (HKEPD) through Hong Kong territory. The experimental comparison between the SVM and the classical radial basis function (RBF) network demonstrates that the SVM is superior to conventional RBF in predicting air quality parameters with different time series. The variability of performance regarding to the free parameters of SVM is also examined and discussed.

1

Gardner, MW, Dorling, SR, "Artificial neural networks (the multilayer feed-forward neural networks) - a review of applications in the atmospheric science", Atmospheric Environment, 30(14/15), 2627-2636, 1998. doi:10.1016/S1352-2310(97)00447-0

2

Vapnik V, "The Nature of Statistical Learning Theory". New York, Springer- Verlag, 1995.

3

Mukherjee S, Osuna E, & Girosi F, "Nonlinear prediction of chaotic time series using a supoort vector machine", Proc. NNSP, 511-520, 1997. doi:10.1109/NNSP.1997.622433

4

Müller KR, Smola AJ, Rätsch G, Schölkopf B, Kohlmorgen J, Vapnik V, "Predicting Time Series with Support Vector machines", Proc. ICANN, 999-1004, 1997. doi:10.1007/BFb0020283

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £82 +P&P)