Forecasting Ozone Levels and Analyzing Their Dynamics by a Bayesian Multilayer Perceptron Model for Two Air-Monitoring Sites in Hong Kong

ABSTRACT

Tropospheric ozone (O₃) has adverse effects on human heath and vegetation. Forecasting its daily maximum level and assessing the factors that influence its dynamics are of great importance to Hong Kong and similar metropolitans in the world. In this article, we simulate the daily maximum O₃ level in Hong Kong by applying the multilayer perceptron (MLP) model trained with the automatic relevance determination (ARD) method in a Bayesian evidence framework. The proposed model is named the MLP-ARD. By using the ARD method, the O₃ influential factors, which are the model's input variables, can be ranked according to their relative importance in regard to the model's output variable, that is, the daily maximum O₃ level. The formation and transportation mechanism of O₃ for two selected air-monitoring sites can be grossly explained by the ranking information. Compared with the MLP model trained by the Levenberg–Marquardt algorithm, the predictive performance of the MLP-ARD for the aforementioned air-monitoring sites is more reliable and accurate in both episode and non-episode periods.

Key Words:

• Bayesian evidence framework
• maximum likelihood
• multilayer perception model
• daily maximum Ozone level
• ozone episode
• relevance determination method.

ACKNOWLEDGMENTS

The work described in this article was partially supported by grants from the Research Grants Council of the Hong Kong Special Administrative Region (HKSAR), China [Projects No. CityU 1185/05E and CityU 1001–PPR20051].

Notes

¹Over-fitting occurs when a MLP is too complex or the number of training data is too small. Such MLPs may fit the noise, not just the signal in a training dataset during the training stage and its performance usually deteriorates on unseen data.