APPLICATION OF NEURAL NETWORKS FOR ATMOSPHERIC POLLUTION FORECASTING

Abstract

A large number of hazardous emissions from industrial production is an environmental problem for the world cities. In the field of environmental engineering, the study of air quality and the prediction of changes in concentrations of harmful substances will make it possible to develop the right strategies for sustainable development. The paper presents the results of the development and research of the applicability of neural network modeling models for forecasting and distribution of concentrations of emissions into the atmosphere by the example of Ust-Kamenogorsk, Kazakhstan. For forecasting, the authors used an RNN network of the long short-term memory (LSTM) type, which is well adapted to learning using the tasks of classification, processing and forecasting of time series in cases when the time periods between events have different gaps. An LSTM model with 3 hidden layers and 1 output neuron in the output layer was determined to predict contamination. To determine the effectiveness of the neural network under study, the average absolute error was used as a function of losses. The authors have developed a system for modeling the process of predicting the pollution of harmful substances in the atmospheric air using data from stationary monitoring points.