Ground-level ozone forecasting using data-driven methods

Accurate site-specific forecasting of hourly ground-level ozone concentrations is a key issue in air quality research nowadays due to increase of smog pollution problem. This paper investigates three emergent data-driven methods to address the complex nonlinear relationships between ozone and meteorological variables in Hamilton (Ontario, Canada). Three dynamic neural networks with different structures: a time-lagged feed-forward network, a recurrent neural network neural network, and a Bayesian neural network models are investigated. The results suggest that the three models are effective forecasting tools and outperform the commonly used multilayer perceptron and hence can be applicable for short-term forecasting of ozone level. Overall, the Bayesian neural network model's capability of providing prediction with uncertainty estimate in the form of confidence intervals and its inherent ability to prevent under-fitting and over-fitting problems have established it as a good alternative to the other data-driven methods.