A Hybrid Framework for Imputing Time Series Air Pollutant Data in Tropical Climates Area

Air pollution poses a significant threat to human health and environmental quality, particularly in densely populated tropical cities where rapid changes in meteorological conditions complicate pollutant dispersion. Accurate forecasting of fine particulate matter (PM2.5) relies on continuous high-resolution time series data, but ambient monitoring often suffers from extensive missing values due to sensor maintenance, communication failures, and extreme weather events. Despite advances in statistical and machine learning imputation methods, it remains unclear how structured hybrid frameworks that integrate seasonal decomposition, multivariate statistical imputation, and recurrent neural networks affect the performance of state-of-the-art sequence models. Here we show a hybrid imputation framework that integrates seasonal-trend decomposition via loess, residual interpolation, and gap classification. For moderate gaps it applies multivariate imputation by chained equations, while for long hydrocarbon gaps it employs a gated recurrent unit with decay, ultimately producing a complete dataset for model training. The proposed Hybrid STL-MICE-GRU framework demonstrated robust performance in restoring data continuity and maintaining the temporal and statistical properties essential for accurate shortterm PM2.5 forecasting. These findings indicate that targeted hybrid imputation not only enhances data quality but also unlocks the full potential of advanced deep learning models for more reliable air quality forecasting, with direct implications for public health decision making and environmental management.