Nonlinear Optimization of Recurrent Neural Networks in the Prediction of Air Quality

Ambient air pollution remains a major threat to public health, particularly in low- and middle-income countries where monitoring infrastructure is often limited. Fine particulate matter (PM2.5) is of particular concern due to its strong association with respiratory and cardiovascular diseases. This study presents a nonlinear optimization framework designed to enhance the predictive performance of Long Short-Term Memory (LSTM) networks in forecasting PM2.5 concentrations. Using a large-scale dataset of more than 506,000 hourly air quality observations collected across Uganda, the proposed model integrates a constrained nonlinear optimization mechanism into the LSTM architecture. This approach reduces overfitting, improves generalization, and provides a theoretically grounded solution to recurrent neural network instability. Experimental results show that the model achieved a Mean Squared Error (MSE) of 0.5968, a Mean Absolute Error (MAE) of 0.5276, and a coefficient of determination (R²) of 0.1875. Feature importance analysis indicated that PM10, carbon dioxide (CO₂), and humidity were the most influential predictors of PM2.5 concentrations. The model successfully captured long-term pollutant trends and seasonal variations, but it underestimated extreme pollution events, suggesting the potential value of ensemble strategies and cost-sensitive learning for improving sensitivity to peaks. Despite these limitations, the findings underscore the viability of integrating nonlinear optimization techniques within deep learning frameworks to achieve more robust and stable predictions. Importantly, this work contributes to advancing air quality forecasting in resource-constrained environments and provides a methodological foundation for future studies that aim to incorporate broader environmental and socio-economic factors.