Time-Series Forecasting of Urban Air Pollutant Levels Using Deep Learning Models

Accurate forecasting of urban air pollutants is critical for public health management and regulatory compliance.While machine learning models are widely used, their performance is highly dependent on the robustness of the underlying methodological framework.This study proposes and validates a comprehensive, end-to-end framework for forecasting hourly Carbon Monoxide (CO) concentrations using the canonical UCI Air Quality dataset.The methodology integrates context-aware data imputation, an advanced feature engineering pipeline (incorporating temporal, cyclical, autoregressive, and rolling-window features), and a rigorous comparative evaluation of fourteen distinct machine learning models.Crucially, all models are validated using a 5-fold TimeSeriesSplit cross-validation protocol to ensure temporal data integrity and prevent lookahead bias.The results demonstrate the clear superiority of ensemble methods, with an optimized XGBoost model emerging as the top performer, achieving an R-squared score of 0.9216 and a Root Mean Squared Error of 0.3824 mg/m.Feature importance analysis revealed that the model's predictions were overwhelmingly driven by non-methanic hydrocarbon (NMHC) sensor readings and their engineered non-linear terms, confirming the model learned scientifically sound relationships.This study validates a synergistic framework where advanced feature engineering paired with powerful ensemble models provides a new benchmark for accuracy and offers a viable template for developing reliable, real-world air quality forecasting systems.