Speaker
Description
Air quality profoundly impacts human health, particularly in rapidly urbanizing regions like South Africa, where pollution from industrial and vehicular sources poses significant risks. Recent reviews highlight Long Short-Term Memory (LSTM) networks as superior for modeling temporal dependencies in air quality data, outperforming traditional methods in pollution forecasting. Building on these advancements, this study applies LSTM to real-time data from multiple South African Consortium for Air Quality Monitoring (SACAQM) sites across Gauteng and beyond.
We preprocess multivariate time-series data including PM$_{2.5}$, NO$_2$, SO$_2$, and meteorological variables from diverse monitoring stations, addressing challenges like missing values and sensor drift through imputation and normalization. A stacked LSTM architecture with attention mechanisms captures both short and long term pollution patterns, trained on historical SACAQM datasets and validated via cross-site temporal splitting.
Results demonstrate superior performance, with RMSE reductions of 15--25\% over baselines like ARIMA and standard RNNs, and $R^2$ scores exceeding 0.92 for multi-step ahead predictions. Spatial visualizations reveal urban-rural gradients and episodic events (e.g., winter inversions), offering physics-based insights into pollutant transport.
This work advances LSTM applications for localized air quality management, providing actionable forecasts for public health alerts and policy in South Africa. Future extensions include hybrid physics-ML models for enhanced interpretability.
| Apply for student award at which level: | MSc |
|---|---|
| Consent on use of personal information: Abstract Submission | Yes, I ACCEPT |