|
Comparative Study of Hybrid Machine Learning and TimeSeries Models for Multi-Horizon Air Quality Forecasting:The Case of Ba Ria – Vung Tau, Vietnam |
|---|---|
| รหัสดีโอไอ | |
| Creator | Ho Minh Dung |
| Title | Comparative Study of Hybrid Machine Learning and TimeSeries Models for Multi-Horizon Air Quality Forecasting:The Case of Ba Ria – Vung Tau, Vietnam |
| Contributor | Nguyen Hien Than |
| Publisher | Thai Society of Higher Education Institutes on Environment |
| Publication Year | 2569 |
| Journal Title | EnvironmentAsia |
| Journal Vol. | 19 |
| Journal No. | 3 |
| Page no. | 110-128 |
| Keyword | Forecast, AQI, Gieng Nuoc intersection monitoring station, Neural network, LSTM |
| URL Website | http://www.tshe.org/ea/index.html |
| Website title | EnvironmentAsia |
| ISSN | 1906-1714 |
| Abstract | Ba Ria-Vung Tau is a rapidly developing economic hub in southern Vietnam with strongrobust growth in the industry, seaports, and transportation. However, this development hasexerted significant pressure on air quality, increasing pollutant concentrations and posingrisks to public health and ecosystems. Although various machine learning and hybrid modelshave been investigated for air quality forecasting, there remains a paucity of studies havecomprehensively compared the predictive performance of hybrid deep learning and traditionaltime-series models using normalized Air Quality Index (AQI) data across multiple forecastinghorizons. To address this gap, this study applied Long Short-Term Memory (LSTM),Long Short-Term Memory–Moving Average (LSTM-MA), Long Short-Term Memory–AutoRegression Integrated Moving Average (LSTM-ARIMA), Nonlinear Autoregressive NeuralNetwork (NAR), Moving Average (MA), and Auto Regression Integrated Moving Average(ARIMA) models to data-normalized AQI data collected at the Gieng Nuoc intersection airquality monitoring station (Ba Ria-Vung Tau) during 2020–2024, and systematically comparedtheir predictive performance for 7-day, 15-day, and 30-day forecasting horizons. Results showthat hybrid and deep learning models consistently outperform traditional time-series approaches,with the hybrid LSTM-MA model demonstrating the best overall performance; at the 15-dayforecasting horizon. The RMSE values are 0.025, 0.068, and 0.109 for the training, testing,and forecasting sets, respectively. While the MAPE values obtained for the validation andforecasting sets are 1.73% and 17.72%, respectively. These findings highlight the comparativeadvantage of hybrid models and support their application in early-warning systems andair-quality management.Keywords: Forecast; AQI; Gi |
