Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Transactions on Ecology and the Environment vol 6, © 1995 WIT Press, www.witpress.com, ISSN 1743-3541 Neural networks - a new mathematical tool for air pollution modelling M. Boznar, P. Mlakar Jozef Stefan Institute, Jamova 39, SI-61000, Ljubljana, Slovenia ABSTRACT Modelling is a basis for better understanding and prevention of air pollution. Many different mathematical techniques have been used in this field. In this paper a new tool for air pollution modelling neural networks - will be described. Its capabilities and disadvantages will be shown in the cases of wind forecasting and reconstruction, and SC>2 pollution forecasting. INTRODUCTION Air pollution is one of the biggest problems of industrialised areas. In order to better understand this phenomena, several different modelling techniques have been applied in this field that are very well known. Research in mathematical fields and in many other modern fields of science also gave some rather new tools that can be usefully employed in the field of air pollution modelling. One such tool, that so far was very rarely used in air pollution modelling, is that of neural networks, which are widely used in other fields such as pattern recognition, speech recognition and synthesis, financial forecasting, etc. The performance of neural network based models will be shown with examples of wind reconstruction and SO2 pollution forecasting in the Sostanj basin (complex orography) in Slovenia. The Sostanj Thermal Power Plant Transactions on Ecology and the Environment vol 6, © 1995 WIT Press, www.witpress.com, ISSN 1743-3541 260 Air Pollution Theory and Simulation (STPP) is a single large source of SC>2 pollution (see Figure 1). A modern Environmental Information System (EIS STPP) is in operation around the plant, providing a huge data base suitable for modelling (e.g. Lesjak^). Figure 1: Map of the Sostanj basin with Thermal Power Plant and automatic measuring stations. ARTIFICIAL NEURAL NETWORKS Artificial neural networks are mathematical structures that simulate the structure and behaviour of the human brain. They are capable of classification, of learning from the multitude of learning samples without knowing any physical laws and equations, of generalisation etc. A neuron (see Figure 2) is the basic element of the neural network. It receives information from the input links and produces the output. Combining many neurons into a structure (see Figure 3) gives a neural network with learning and generalisation capabilities. Neural networks differ from each other by their structure (number and organisation of neurons), and type of transfer function used in the neurons. In the process of learning the neural network adjusts its interconnection weights (coy) in order to minimise the error of predictions over the known training set of data. When adjustment is made, the performance of the model can be tested using data that were not included in the training set. Transactions on Ecology and the Environment vol 6, © 1995 WIT Press, www.witpress.com, ISSN 1743-3541 Air Pollution Theory and Simulation 261 OUTPUT INPUTS Figure 2: Typical neuron INPUT LAYER _ HIDDEN LAYER OUTPUT LAYER NEURON BASIC ELEMENT STRUCTURE Figure 3: Typical structure of an artificial neural network For our modelling we used a Perceptron neural network (e.g. that of Lawrence*) with one hidden layer, and a sigmoid non-linear transfer function that gives remarkable training and generalising capabilities. GROUND LEVEL WIND FORECASTING Meteorological data, especially wind measurements are crucial for the air pollution forecasting. If wind data were known in advance (for example, a half an hour interval in advance), it would be easier to predict plume movements. The first application was a model that predicts wind measurements at the Graska gora station (measurements at 10m). Separate models were built Transactions on Ecology and the Environment vol 6, © 1995 WIT Press, www.witpress.com, ISSN 1743-3541 262 Air Pollution Theory and Simulation 21 31 case n u m b e r Figure 4: Ground level wind direction prediction at Graska gora station. 301 401 case number predicted measured] fffH+H-fH ! 1 1 1 ! 1 1 1 i M 1 1 1 ! fti 41 61 case number Figure 5: Ground station. level wind speed prediction 81 at Graska gora Transactions on Ecology and the Environment vol 6, © 1995 WIT Press, www.witpress.com, ISSN 1743-3541 Air Pollution Theory and Simulation 263 for wind speed and wind direction prediction. The inputs were the wind speed and wind direction measurements at the stations of Graska gora and Veliki vrh for the current half an hour interval and for one measuring interval in the past for wind speed. For the prediction of wind direction, measurements from the station at Zavodnje were added. The results are shown in Figure 4 and Figure 5. Wind direction prediction seems to be a difficult problem, because the results were not good at all. The wind speed prediction model performs better, but the tendency to use the current measurement as the predicted value can be seen from the detailed graph (Figure 5). RECONSTRUCTION OF SODAR MEASUREMENTS Another useful application would be a model for reconstruction Figure 6: Reconstruction of SODAR data (error of wind direction at 50m) (not forecasting) of missing SODAR data, because wind profiles are essential for advance numerical models and usually only interpolations are used when there are no measurements available at a particular level. For the Sostanj basin SODAR measurements are available for one month in the spring of 1991, when a measuring campaign was performed there (e.g. Elisei et. aU). As the input for the model, ground level wind measurements at Transactions on Ecology and the Environment vol 6, © 1995 WIT Press, www.witpress.com, ISSN 1743-3541 264 Air Pollution Theory and Simulation different stations were taken (without SODAR measurements!). The outputs were wind speed and direction at a particular SODAR level. The results, as shown in Figure 6 and Figure 7, are acceptable. ° I 8 8 I Figure 7: Reconstruction of SODAR data (wind direction; solid line - measured; dotted line - reconstructed) PREDICTION OF AMBIENT SO2 CONCENTRATIONS Our most successful application is a neural network - based model for ambient SO2 concentration prediction. The results were reported in the literature (e.g. Boznar^), and at conferences (e.g. Mlakar^, Mlakar&). The model predicts ambient SO2 concentrations at the Zavodnje station for half an hour in advance. The inputs are meteorological measurements (without SODAR data), ambient pollutant concentration measurements at several stations and TPP emission data. The results were very good. Some are shown in Figure 8. Detailed analysis of the predictions shows that the model predicts very good high peaks of concentration during thermal inversion situations that are very difficult even to reconstruct with other types of models. Transactions on Ecology and the Environment vol 6, © 1995 WIT Press, www.witpress.com, ISSN 1743-3541 Air Pollution Theory and Simulation 265 Testing patterns measured calculated O <-> m <N O i j 4. o 4^ā^āiāfā! 0 % 50 case number % 100 Figure 8: Prediction of ambient SO2 concentrations for half an hour in advance at Zavodnje station. CONCLUSIONS Three different applications of neural networks were shown to encourage further research in this field. The most difficult problem was prediction of wind speed and direction. Slightly better results were obtained in reconstruction of SODAR vertical wind profiles. The most suitable application for neural network based models seems to be the prediction of pollutant concentrations. For all these applications more effort should be put into finding proper inputs for the model (input features), and finding information-rich sets of learning and testing data. REFERENCES 1. J. Lawrence Introduction to Neural Networks California Scientific Software, Grass Valley, 1991. 2. M. Boznar, M. Lesjak, P. Mlakar, A neural network-based method for short-term predictions of ambient SO2 concentrations in highly polluted industrial areas of complex terrain, Atmospheric Environment, Vol.27B, pp. 221 - 230, 1993 Transactions on Ecology and the Environment vol 6, © 1995 WIT Press, www.witpress.com, ISSN 1743-3541 266 Air Pollution Theory and Simulation 3. M. Lesjak, B. Diallo, P. Mlakar, Z. Rupnik, J. Snajder and B. Paradiz., Computerised ecological monitoring system for the Sostanj thermal power plant, in Man and his ecosystem, pp. 3-31 to 3-38, Proc. 8th World Clean Air Congress, Hague, The Netherlands, 1989. 4. Elisei et. al., Experimental Campaign for the Environmental Impact Evaluation of Sostanj Thermal Power Plant (1992), Jozef Stefan Institute, ENEL/DSR/CRTN (Milano), CISE (Milano) Progress Report, (1992) 5. Primoz Mlakar, Marija Boznar, Martin Lesjak, Neural networks predict pollution, Air Pollution Modelling and Its Application X, ed. Gryning, Millan, pp. 659 to 660, Plenum Press, New York, 1994 6. P. Mlakar, M. Boznar, Short-term air pollution prediction on the basis of artificial neural networks, pp. 545 to 552, Air Pollution II Volume 1: Computer Simulation, Computational Mechanics Publications, Southampton, Boston, 1994