Time series analysis and prediction of PM2.5 pollution concentration in Igdir province with deep learning models

Abstract

Air pollution is a global problem that causes serious environmental and health problems, especially in regions where industrialization and urbanization are intense. Igdir province, located in the eastern part of Türkiye, is one of the regions where air pollution is intensely observed due to its geographical structure and meteorological characteristics. This study evaluated deep learning models for predicting PM2.5 levels using data from national monitoring networks, meteorological services, and NASA POWER. Preprocessing included interpolation, outlier correction, and min–max normalization. LSTM, GRU, Bi-LSTM, Bi-GRU, CNN-LSTM, and CNN-GRU models were tested across 8, 24, and 72 h windows. The GRU model achieved the best performance in short-term (8 h) predictions with MAE=9.93 and R${}^2$=0.944 values. The LSTM model reached the best predictive performance for the 24 h window with MAE=9.65 and R${}^2$=0.949, while for the 72 h window, the BiLSTM model outperformed the others. In terms of predicting peak values, the CNN-LSTM model stood out, achieving RMSE=28.16, R${}^2$=0.792, and MAE=22.45 in the 8 h window. These findings highlight deep learning’s efficacy for air pollution forecasting and decision support.