Multivariate Forecasting of Multiple Pollutants with Representative Deep Learning Architectures

This study introduces new deep learning methodolo-gies to improve multivariate multi-pollutant forecasting. Leveraging real data from three cities acquired in different time frames, we develop five regression models based on representative deep modules to model spatial and temporal information to predict concentrations of PM<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>.<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">5</inf>, PM<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10</inf>, NO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>, CO, SO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>, and O<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</inf>. We evaluate the performance and generalizability of the models trained with and without data augmentation. In addition, we analyze the influence of historical data volume on predictive accuracy. Experimental results show that LSTM-based architectures achieve better performance than our other proposals and methods in the state of the art. Our research aims to advance proactive air quality management by fostering healthier and more sustainable urban environments.