Prediction of particulate matter based on meteorological parameters using deep learning : an analysis of their relationship

Abstract

The primary aim of this thesis was to examine the utilization of deep learning algorithms in predicting particulate matter concentrations based on meteorological parameters in Sarajevo, the capital city of Bosnia and Herzegovina. Furthermore, correlations between particulate matter and meteorological parameters were investigated. The data set was provided by the Hydrometeorological Institute of Bosnia and Herzegovina for the period from the beginning of 2022 till the end of 2023. The data set consisted of hourly measured values for particulate matter concentrations and meteorological parameters during the mentioned period. The predictions were made using three distinct models: feed-forward neural networks (FFNN), recurrent neural networks (RNN), and long short-term memory networks (LSTM). To explore the correlation between meteorological parameters, particulate matter with temperature inversion, the dataset was separated into two separate cases. In the first case, all accessible parameters were considered, including temperature data. In the second case, temperature data were replaced with temperature inversion data to examine the effects of this particular meteorological phenomenon and its correlations with other parameters. Models with the best performances were feed-forward neural networks. The capacity to precisely estimate pollutant concentrations has the potential to minimize the number of chronic patients in the future and improve the timely and appropriate reaction of competent institutions.