PMFORECAST: leveraging temporal LSTM to deliver in situ air quality predictions

The physical and chemical properties of atmospheric aerosol particles are crucial in influencing global climate and ecosystem processes. Given the numerous studies highlighting adverse health effects from exposure to aerosol particulates, particularly Particulate Matter (PM), effective air quality management strategies are under consideration  (Annesi-Maesano et al. Eur Respir Soc 29(3):428-431. 2007). Herein, we introduce a predictive model-PMFORECAST-employing a self-adaptive long short-term memory (LSTM) architecture to predict PM $$ values in the real atmosphere. Specifically, we explore adopting a LSTM model to better benefit from temporal dimensions. PMFORECAST is strategically designed with four key phases: preprocessing, temporal attention, prediction horizon, and LSTM layers. By leveraging LSTM's significant predictive ability in time-series data, the inclusion of temporal attention enhances the model's specificity. Temporal dynamics modeling entails generating insights over time, utilizing temporal attention to extract essential characteristics from historical air pollutant concentrations, with the flexibility to adjust the historical data according to the forecasting period. To assess PMFORECAST, we consider measurements collected from the QAMELEO network, a sparse network of air-quality micro-stations deployed in Dijon, France. The self-adaptive capabilities of PMFORECAST allow the model to be dynamically updated, evaluating its performance and continuously tuning hyper-parameters based on the latest data trends. Our empirical evaluation reports that PMFORECAST outperforms the state of the art, achieving notable accuracy in both short-term and long-term predictions. The PMFORECAST deployment at scale can serve as a valuable tool for proactive decision-making and targeted interventions to mitigate the health risks associated with air pollution.