Akademik Veri Yönetim Sistemi
Ozone: Science and Engineering, cilt.47, sa.5, ss.575-582, 2025 (SCI-Expanded)
In the face of growing global concern about air pollution and its adverse impacts on health and well-being, air quality forecasting has become a major focus of research. This study investigates the application of Long Short-Term Memory (LSTM) neural network models for the prediction of ozone concentrations in different time windows (7, 15 and 30 days). The LSTM models were configured with two LSTM layers and two dense layers, varying the number of neurons between 50 and 200. Predictor variables included ‘Month,’ CO_ppb,’ ‘SO2_ugm3,’ ‘temperature,’ ‘humidity’ and ‘PM2.5_ugm3,’ while the output variable was ‘O3_ppb.’ The results showed that the LSTM model with 200 LSTM units and 50 dense units had the best performance, presenting a mean absolute error (MAE) of 1.49 ppb, an RMSE of 1.91 ppb and a coefficient of determination (R2) of 0.94 for 7-day forecasts. These results outperformed traditional models such as ARMA and linear regression, highlighting the effectiveness of LSTM in capturing complex patterns in ozone time series. The research suggests that future investigations could explore hybrid techniques to further improve the accuracy of predictions.