Nitrogen oxides (NOX=NO+NO2) are emitted into the urban atmosphere primarily from vehicle exhausts. Primary NOX emissions are mostly in the form of nitric oxide (NO) which then reacts with ozone (O3) to form nitrogren dioxide (NO2). Much work has been carried out to determine the factors which control the NOX and NO2 concentrations in order to enable the development of tools to aid in the forecasting of pollutant concentrations.
One approach to predict future concentrations is to use detailed atmospheric diffusion models. Such models aim to resolve the underlying physical and chemical equations controlling pollutant concentrations and therefore require detailed emissions data and meteorological fields for the region of interest.
The second approach is to devise statistical models which attempt to determine the underlying relationship between sets of input data (predictors) and targets (predictands). Regression modelling is an example of such a statistical approach and has been applied to air quality modelling and prediction by a number of people (Shi & Harrison, 1997; Ziomass et al., 1995). One limitation of linear regression models is that they will underperform when used to model non-linear systems (Gardner & Dorling, in press). The relationship between NOX, NO2 and meteorology is complex and extremely non-linear.
In this work multilayer perceptron (MLP) neural networks are used to model and predict hourly urban NOX and NO2 concentrations from readily observable local meteorological data. MLP neural networks are capable of modelling highly non-linear relationships and can be trained to accurately generalise when presented with new, unseen data. MLP neural networks learn to model a relationship during a supervised training procedure, when they are repeatedly presented with series of input and associated output data. In the case of modelling polutant concentrations the input data will consist of measurements of meteorological conditions and the output will be the pollutant concentration. To use the trained MLP neural network for prediction involves presenting the network with a set of forecast meteorological data. For this reason it is important to use meteorological data that are readily forecasted.
Results from this work show that the MLP models can account for 60% of the variability of the hourly NOX and NO2 concentrations when used to predict one year of concentrations. This is considerably better than previous linear regression and auto-regressive models (see for example Shi & Harrison (1997)). It is also demonstrated that the MLP models are resolving the diurnal pattern of source strength without any external guidance. Work is being undertaken to assess the extent to which the MLP model predictions are degraded when made using forecast meteorological data. Current work includes the construction of a neural network ``trajectory box model'' ozone forecasting scheme. This work will build on encouraging results which have demonstrated the versatility and consistent performance of MLP neural network air quality models in a range of urban and rural settings for both primary and secondary pollutants throughout the United Kingdom.