Abstract
A neural-networks-based method (NN-AirPol) for air pollution evaluation and control was proposed. The most popular neural networks, the backpropagation algorithms, were used to model the relationship between local meteorological data and air pollution indicator concentrations like SO2, PM10, etc. NN-AirPol is a simple, feasible and versatile method to predict the future air pollution in urban cities due to the combination of excessive use of low-quality fuels and unfavourable meteorological conditions that could result in catastrophic effects on human health. After backpropagation training, the neural network evaluates the forecasted concentration of critical air pollution indicators. Depending on their concentration values, relevant episode warnings and actions are activated. For illustrating and validating of NN-AirPol method, a case study based on air pollution parameters and data from the Turkish State Meteorological Organisation, Istanbul-Göztepe station, was carried out. The proposed method is believed to be crucial for rapidly developing countries like Turkey. Therefore, as air pollution indicators (neural network outputs), sulphur dioxide and inhalable particulate materials were considered. As the best among ten backpropagation algorithms, the BFGS algorithm (Quasi-Newton algorithm) was selected. The optimal architecture of NN-AirPol neural network was determined. A comparison of the method with regression and perceptron models showed that it gives significantly better performance. The proposed method can predict many air pollutants in a metropolitan city like Istanbul and consequently offer the appropriate episode warning signals and the relevant actions to be taken by the government or the public to reduce that particular pollutant to a non-harmful level. In summary, the NN-AirPol method showed the following advantages: Significantly better performance (much lower error rate) than linear neural-network-based methods like the multi-layer perceptron and statistical methods like the linear regression model. NN-AirPol models better non-linear dependencies characterising the air pollution data. Predictions of SO2 and PM10 generated by NN-AirPol were 2-13 times and 2-5 times more accurate than other linear methods, respectively. Long-term prediction of daily concentration of the criteria pollutants over a city and short-term prediction based on hourly data for prediction of next hour or half an hour air pollutants concentration. Predicting (not measuring) the emission concentrations one day before it happens is enough time to take proper actions for preventing the production of pollutants before they are emitted. In the case of using predicting meteorological data, three-day pollution parameters could be predicted, on the condition that predicted meteorological parameters should be obtained. Connection with air pollution warning system providing temporal information about the episode of pollutants. The 'red and yellow episode warnings and actions' represent the exceedance of the relevant threshold values. The NN-AirPol method can have the following further developments: Application and validation of NN-AirPol with further air pollution indicators like nitrogen compounds, carbon monoxide (CO), ozone (O3), etc. The method was applied to predict the daily concentration of the criteria pollutants over Istanbul city, which is a long period. It would be good to apply it to hourly data and predict air pollutant concentration at the next hour or the next half an hour. A software tool implementing NN-AirPol method will assure its easy practical application. This tool could be connected with an eco-warning system.