Neural Network Modelling of Present and Future Urban PM10 Concentrations based on Measurement Results

Thesis of Keng Been Ang
Universität Stuttgart, 2010

The problem of air pollution is a frequently recurring situation and its management has considerable social and economic effects. On one hand, air pollution control is necessary to prevent the situation from worsening in the long run. On the other hand, forecasting of air quality in days in advance is also necessary in order to adopt preventive actions during episodes of airborne pollutions.

In the past decade, neural network models have become an efficient tool for establishing both temporal and spatial characteristics of ambient air quality in the field of air pollution. Neural network models are capable of learning 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 ambient air pollutant concentrations, the input data could consist of meteorological or air quality data from measurements, and the outputs would be the air pollutant concentrations.

In this dissertation, the objectives are to realise and to evaluate two air quality neural network models that, correlating the air quality data with the meteorological information, are able to simulate daily urban PM10 concentrations in Stuttgart. For that, two models are developed for PM10 Nowcasting and PM10 Forecasting.

To conduct a thorough and insightful evaluation on the modelled PM10 concentrations, several performance indices which were used for both the PM10 Nowcasting and PM10 Forecasting models included the fractional bias, the index of agreement, the squared correlation coefficient, the mean absolute error, the mean bias error and the root mean square error. For the PM10 Forecasting model, additional performance indices to evaluate the correct number of PM10 exceedances were considered. For both models, thorough analyses of the error residuals and quantile-quantile plots were performed for the identification of possible outliers and for the better understanding in the patterns across the two sets of univariate modelled and measured data.

From 15.11.2006 to 18.03.2007, intensive street sweeping was conducted along the paved roadway of Stuttgart Neckartor as an urban PM abatement strategy. Based on results from single particle analyses and measurements of PM and NOX concentrations, reductions in ambient PM10 concentrations could be suggested. However, an exact quantitative evaluation on the effectiveness of street sweeping on ambient PM10 was complicated by the possible influence of different meteorological conditions and other unknown factors during sweeping and non-sweeping days. With the neural network approach, these influencing meteorological conditions could be parameterised as functions to PM10 concentrations. The PM10 Nowcasting model was thus developed as a tool to complement the results of the past measurements. The aim of the developed model is to Nowcast the original state of PM10 concentrations at Neckartor, assuming that no street sweeping activities took place during the sweeping periods. Any effect of street sweeping could then be suggested by any differences between the modelled PM10 concentrations and the corresponding measured PM10 values. Through extensive statistical evaluation on the performance of the developed model, it was capable of accurately simulating past PM10 concentrations from January 2004 to October 2006. For the next step, the suitability of the developed model for operational use was then evaluated for the modelling of PM10 concentrations at the Neckartor site on the 41 days with street sweeping. Although results from linear regression analysis between the modelled PM10 concentrations against the measured values showed that the measured PM10 values were approximately 4 % lower than the modelled values, trends of lower PM10 concentrations were not observable during all sweeping periods. Interesting, this reduction trend from the modelling results was in accordance to the measurement results.

The PM10 Forecasting model was developed to forecast the daily PM10 concentrations in one, two and three days in advance for two urban sites of different characteristics in Stuttgart. The first site represented the heavily trafficked site Neckartor, and the second site represented the urban background site Bad Cannstatt. The input parameters were on the one hand measurement data from the two ambient air monitoring stations at Neckartor and Bad Cannstatt. On the other hand common forecasted weather parameters up to three days in advance, which were obtained from a Numerical Mesoscale Model, were included. The overall model’s results illustrate a possibility of effective use on the operational level for performing future PM10 forecasts up to three days at both the traffic and urban background sites. However, in real-time forecasting conditions, a compromise in performance should be expected, due to the possibility of less accurate meteorological forecasts. Therefore, a prerequisite for the successful implementation for PM10 forecasting is the availability of high quality meteorological forecasts, as the model performs according to the accuracy of these parameters.

Both the PM10 Nowcasting and PM10 Forecasting models encountered difficulties in accurately simulating PM10 concentrations during several distinct PM10 episodes. From the mathematical aspect, the underpredicting behaviours of both models during episodic events verifies the general assumption that neural network models will fail to extrapolate on data which have not been presented during the training procedure. From the scientific aspect, the underpredicting behaviours of the models could be attributed to the additional loads of PM from episodic events, whose presence could not be accurately modelled by the input parameters. The three most probable types of PM10 episodes are the extreme wintertime inversion-induced PM10 episodes, recreational PM10 episodes and regional and long-range PM10 transport.

A general conclusion is that neural network models can be useful and fairly accurate tools of assessment in PM10 concentrations in urban areas. However neural network models have inherent limitations. In this dissertation, the main limitation is that both PM10 Nowcasting and PM10 Forecasting models are strictly site-specific. Nevertheless, the general approach can be followed, especially in the case of neural networks, where a number of key decisions on their formulation, topology and operating parameters are necessary for the accurate simulation of PM10 concentrations.
To the top of the page