With the rapid growth in the availability of data and computational technologies, multiple machine learning frameworks have been proposed for forecasting air pollution. However, the feasibility of these complex approaches has seldom been verified in developing countries, which generally suffer from heavy air pollution. To forecast PM2.5 concentrations over different time intervals, we implemented three machine learning approaches: multiple additive regression trees (MART), a deep feedforward neural network (DFNN) and a new hybrid model based on long short-term memory (LSTM). By capturing temporal dependencies in the time series data, the LSTM model achieved the best results, with RMSE = 8.91 µg m–3 and MAE = 6.21 µg m–3. It also explained 80% of the variability (R2 = 0.8) in the PM2.5 concentrations and predicted 75% of the pollution levels, proving that this methodology can be effective for forecasting and controlling air pollution.