Machine Learning Approaches to Develop Weather Normalized Models for Urban Air Quality
Urban air quality is a growing concern worldwide, with a significant portion of the global population living in areas where air pollution exceeds recommended thresholds. The impact of air pollutants, such as particulate matter (PM) and gases like SO2, CO, NO2, O3, can have serious implications for human health. In response, many countries have implemented regulations and interventions to reduce these effects.
One key aspect of assessing air quality is understanding the relationship between pollutant concentrations and various factors, including weather conditions and human activities. Traditional models, however, struggle to accurately capture the complex and non-linear nature of this relationship. This is where machine learning (ML) and deep learning (DL) techniques come into play.
The primary objective of a recent study was to explore the power of ML and DL techniques in developing weather normalized models (WNMs) and improving their accuracy. The researchers also aimed to use these enhanced models to evaluate the impact of events on air quality. These ML/DL-based WNMs were found to be valuable tools for conducting exploratory data analysis (EDA) and identifying correlations between meteorological and temporal features and air pollutant concentrations.
The study compared various DL architectures, such as LSTM, RNN, BiRNN, CNN, and GRU, to develop the WNMs. Among these architectures, LSTM-based methods demonstrated superior results. The researchers discovered that their DL-based WNMs could effectively capture the correlations between meteorological variables and five criteria contaminants (SO2, CO, NO2, O3, and PM2.5) using the SHAP library.
The researchers also applied these WNMs to assess air quality changes during the COVID-19 lockdown periods in Ecuador. Existing normalized models typically operate based on average or consistent weather conditions and are not designed for predicting pollution peaks, which often lack discernible patterns. To address this limitation, the study enhanced the WNMs specifically to perform well during daily concentration peak conditions. Supervised learning techniques, including Ensemble Deep Learning methods, were used to distinguish between daily peak and non-peak pollutant concentrations.
It is important to note that this method may introduce potential bias when selecting non-peak values. Nevertheless, the results of this research contribute to our understanding of the correlations between meteorological and temporal features and daily concentration peaks of air pollutants, a critical step toward developing effective strategies to mitigate air pollution.
This study highlights the potential of ML and DL techniques in improving weather normalized models for urban air quality. By accurately capturing the complex relationships between pollutant concentrations and various factors, these models can provide valuable insights for policymakers and urban planners. With further research and development, ML and DL have the potential to revolutionize our understanding of air pollution and help create healthier and more sustainable cities.
Source: [Link to the original article]
