Forecasting the degree of air pollution in the industrial region

Abstract

The article considers the possibility of using artificial neural networks to predict the degree of atmospheric air pollution in the industrial region. The neural network approach is very popular and quite effective in solving forecasting problems. Neural networks make it possible to model linear dependencies in the case of a large number of variables. The task of forecasting with the help of NN consists in constructing an optimal NN based on the initial data, teaching it various algorithms, pre-training (if necessary), and making a forecast. To build a high-quality network, several hundred or thousands of observations are sufficient. Network training is a fitting of the model, which is implemented by the network, to the available training data. The developed system works in real time and in a simple and clear form produces quite a high-quality forecast of the pollution level. The system is able to cover all monitoring stations in the industrial region, collecting and analyzing data from at least 100 monitoring stations (as to pollution-every hour, as to the weather – every 15 minutes). The system works 24 hours a day. A multilayer neural network has been developed that allows to predict the level of pollution based on data on the current air quality, current weather conditions, weather forecast, time of day and day of the week, since air quality depends on these characteristics. The forecast is made for each hour, for each station, each pollutant. The forecasting horizon is 6 hours (this is the standard). The accuracy of the prediction for a neural network with a different number of layers was tested. Dependences of the air pollution forecast accuracy on the number of layers of the neural network have been obtained, which show that the maximum accuracy of the forecast is achieved with the use of two external layers and one hidden layer of the neural network. It was determined that the maximum accuracy is achieved with the use of 30 neurons on the hidden layer, which is the optimal solution that ensures a better prediction accuracy. It was confirmed that the creation of one universal neural network, which will predict the level of pollution for any station in the region, does not give an accurate forecast, because stations are in different environments