Siberian Federal University

The authors proposed a new way to predict concentration of nitrogen dioxide, nitrogen oxide, sulfur dioxide and carbon monoxide in the atmosphere, a prolonged exposure to which can impair health and, ultimately, lead to chronic and malignant diseases. The use of neural networks (LSTM) in combination with methods of mathematical modelling made it possible to predict with high accuracy the quantitative values of hazardous pollutants, as well as meteorological conditions. LSTM is currently the most successful type of recurrent neural network that is capable of directly supporting multidimensional sequence prediction problems such as weather forecasting, natural disaster probability and so on.

“Predicting air composition is a challenge for researchers, as it is influenced by many factors: exhaust gases, industrial emissions, coal combustion and dust. Moreover, the speed and nature of the distribution of harmful substances in space are individual for each of them. In their work, the scientists used raw numerical data on the main air pollutants, which were the result of periodic measurements of air quality in 2017–2019, made by atmospheric monitoring stations in Krasnoyarsk,” said Lyudmila Kulagina, assistant professor of the Department of Technosphere and Environmental Safety of Siberian Federal University.

Comparing the concentration of substances with 10 types of meteorological data (temperature, humidity, wind speed, etc.), the authors developed a model architecture for training a neural network. As a result, it was possible to increase the accuracy of the forecast and automate the process of assessing the risks of increasing the level of air pollution.

“The existing models for predicting air pollution without machine learning have serious drawbacks that do not allow them to be widely used for long-term predictions. Artificial intelligence technologies based on LSTM can process not only individual images, but also entire sequences of data (speech, video, and so on), can store information for a given period of time, and selectively change it. Due to the special training system, the built model can track most of the unexpected leaps in the level of air pollution," the researcher explained.

The researchers concluded that the use of different types of weather data can improve the accuracy of air quality predictions for other harmful compounds. The application of the new method will contribute to the development of effective ways to protect the environment and identify sources of pollution.

The journalists of the TASS Science portal posted an article about this study.