Zhang, Rui (2025) Novel Deep Learning-Based Models For Air Quality Prediction: Addressing Non-Stationary And Spatio-Temporal Issues. PhD thesis, Perpustakaan Hamzah Sendut.
![[img]](http://eprints.usm.my/style/images/fileicons/application_pdf.png) |
PDF
Download (483kB) |
Abstract
Accurate air quality monitoring and forecasting are essential for public health, environmental regulation, and urban planning. However, prediction remains challenging due to the inherent non-stationarity of pollutant data, influenced by dynamic emission sources, meteorological variability, and human activities. Complex spatio-temporal dependencies, including lead-lag effects among monitoring stations, further complicate the modeling process. Balancing model accuracy and computational efficiency is crucial. This thesis addresses the challenges of predicting air quality by proposing three innovative models. The first model, named pca-dswt-nlstm, combines principal component analysis (pca), discrete stationary wavelet transform (dswt), and nested long short-term memory (nlstm) for temporal prediction. The second model, named the seasonal-trend spatio-temporal long short-term memory (st2-lstm), is developed for spatio-temporal prediction while considering trend and seasonality. The third model, named spatio-temporal inverted transformer (st-itransformer), is proposed for spatio-temporal prediction with exogenous variables. All three models are rigorously evaluated on real air pollutant data through ablation studies, comparative experiments, and detailed analyses. Results demonstrate that the proposed models consistently outperform benchmark models across various performance metrics, significantly enhancing air quality forecasting capabilities. Specifically, pca-dswt-nlstm excels in pm2.5 prediction with efficient feature extraction and long-term memory. St2-lstmcaptures seasonal and spatial patterns effectively, showing stable and accurate results. St-itransformer delivers high accuracy across multiple stations by modeling spatial lead-lag effects, proving robust for both short-term and long-term forecasts.
| Item Type: | Thesis (PhD) |
|---|---|
| Subjects: | Q Science > QA Mathematics > QA1 Mathematics (General) |
| Divisions: | Pusat Pengajian Sains Matematik (School of Mathematical Sciences) > Thesis |
| Depositing User: | Mr Hasmizar Mansor |
| Date Deposited: | 25 Jun 2026 06:42 |
| Last Modified: | 25 Jun 2026 06:42 |
| URI: | http://eprints.usm.my/id/eprint/64465 |
Actions (login required)
 |
View Item |
