A Complete Air Pollution Monitoring and Prediction Framework

The issue of air pollution is increasingly prominent and represents a significant environmental challenge, particularly in urban areas affected by rising migration rates. Air pollution forecasting is crucial for understanding the mechanisms underlying pollution in a specific region, but analyzing high-dimensional data with spatial and temporal dependencies poses a major challenge for traditional machine learning approaches. Additionally, missing sensor measurements due to malfunctions and connectivity loss have severely limited air pollution forecasting models’ performance and restricted their use in production systems. Although significant efforts have been made in air pollution forecasting, many approaches face challenges in dealing with missing sensor data. Based on past and current research, this paper proposes and evaluates four encoder-decoder architectures with attention for forecasting particulate matter (PM) levels that are location- and season-independent. To handle missing sensor data, this paper also proposes and evaluates two adversarial networks for data augmentation. We conducted experiments to investigate the performance of predictive models with and without augmenting training datasets, and using the proposed adversarial models for data augmentation resulted in superior performance gains. The deep neural architectures developed in this research are general enough for predictive and generative tasks for other pollutants and can be adapted for handling time series data in other domains.