Prediction of Air Pollutant Concentrations by Selected Neural Network Architectures

Abstract

Air pollution currently represents a significant environmental challenge with a direct impact on both public health and ecosystems. One of the aims of this paper is to predict the concentration of particulate matter PM10 and CO2 in the air using machine learning methods, specifically by employing various types of neural networks, including feedforward, convolutional, and recurrent models. The aim is also to compare the performance of these architectures in processing univariate time series with different temporal resolutions (monthly, hourly). The study uses real data from the Slovak Hydrometeorological Institute (SHMÚ) and modelled emissions from the EDGAR database. The models were trained, optimized, and evaluated using standard error metrics. The results showed that Conv1D models achieved the highest accuracy on hourly data of PM10, while feedforward architectures performed best on monthly national data of CO2. The findings confirm the potential of artificial intelligence methods as a tool for environmental monitoring and decision support in air quality protection. Opportunities for extending the models with meteorological variables and applying them in real-time warning systems are also discussed.