Predicting Pollution Before It Strikes: Machine Learning Models for Next-Day PM10 Forecasting Under Extreme Events in Portugal

Air pollution remains a major environmental threat, particularly during extreme events such as wildfires and dust intrusions. This study assesses the ability of two artificial neural network architectures to predict daily PM10 concentrations over the Portuguese Centre region using spatially continuous reanalysis data and meteorological predictors. A Multilayer Perceptron (MLP) and a Deep Learning Long Short-Term Memory (DL-LSTM) model were developed within a retrospective proof-of-concept framework. Data from 2003–2021 were used for training and time-series cross-validation, while 2022 was retained as an independent test year. The target was the daily mean PM10 concentration from the CAMS global reanalysis, and both models were also assessed during the August 2022 episode affecting the Serra da Estrela region. The MLP and DL-LSTM achieved correlation coefficients of 0.93 and 0.92 and RMSE values of 11.72 and 13.58 µg m−3, respectively. Both models obtained a recall of 0.61 for exceedance detection, indicating that approximately 39% of the CAMS reference exceedance occurrences were not detected. The results show similar overall performance for both architectures, with a slight advantage for the MLP, while also highlighting the greater difficulty of predicting the highest PM10 concentrations.