A comparative analysis of deep learning approaches for the prediction of particulate matter concentration

Air pollution is one of the primary factors affecting human health. Poor air quality has adverse physical and mental effects on human health and quality of life. In particular, it is known that particulate matter air pollution leads to serious health problems. Developing reliable models for predicting particulate matter pollution levels (PM2.5 and PM10) will be essential for decision-makers. This study aims to develop high-accuracy artificial intelligence-based prediction models for PM pollution. DL algorithms were run extensive data obtained from air quality monitoring stations, which are environmental facilities. The study consists of two research areas. The first research subject involves an assessment of various LSTM algorithms' applicability in predicting PM10 pollution levels while concurrently exploring the impact of the DPFS process on the predictive accuracy of the LSTM models. In the subsequent phases of the study, three distinct LSTM models are developed: Vanilla, BiDirectional, and Stacked. Experimental results demonstrate that the proposed LSTM models exhibit high prediction performance when the DP-FS process is applied, indicating their usability in predicting hourly PM10 concentrations. The secondary research focus of this study entails a comparative examination of predictive performances across three distinct deep learning algorithms: LSTM, RNN, and GRU, specifically in the context of PM2.5 prediction. The experimental findings highlight a significant superiority of the LSTM+LSTM model over alternative DL algorithms with R2 values of 0.98 and 0.97 for the training and test sets, respectively. In addition, the generalization ability of the model was evaluated with data from nine different districts of Istanbul.