Abstract

Air pollution is one of the main causes of serious health problems such as lung and cardiovascular diseases and increases the risk of other diseases. Therefore, early forecasting of air quality plays an important role in warning the community and helping people proactively take health protection measures. In this study, the LSTM model was selected due to its ability to identify nonlinear relationships in data, suitable for both short-term and long-term forecasting. In order to improve efficiency, the LSTM model was combined with ARIMA to form a hybrid model LSTM-ARIMA, which brought positive forecasting results. The values ​​of CO, NO₂, SO₂ and PM₁₀ parameters at the Gieng Nuoc intersection station (Vung Tau ward, Ho Chi Minh City) were always below the allowable threshold of QCVN 05:2023/BTNMT. The PM_2.5 concentration was mostly at a safe level, only at one point it almost reached the standard threshold. However, the 8-hour average O₃ concentration exceeded the standard at 866 times, and the 1-hour O₃ exceeded the standard at 74 times. The air quality at the Gieng Nuoc intersection monitoring station is mainly in the “Good” and “Average” groups according to the AQI index. The LSTM-ARIMA model gives 1-day forecast results with high accuracy (RMSE = 14.57; MAE = 7.81; MAPE = 16.3%) and maintains stable performance when forecasting is extended to 7 and 14 days, showing potential for application in air quality forecasting. The research results show that the LSTM-ARIMA model has better forecasting performance than single models, and can be scaled up and applied to other observation stations, contributing to the development of flexible forecasting systems that are suitable for the actual observational data conditions in developing cities.