Data-driven one-day-ahead PM₁₀ prediction for Portugal: comparing MLP and LSTM models under extreme fire event

Air pollution is one of the most critical environmental threats to human health and ecosystems, with major socio-economic impacts, and remains the leading environmental health risk in Europe, contributing to hundreds of thousands of premature deaths each year. Accurate one-day-ahead air quality forecasts are therefore essential to support timely mitigation actions and protect vulnerable populations under rapidly evolving atmospheric conditions.

This work develops and evaluates machine learning approaches for next-day prediction of PM₁₀ concentrations in Portugal, focusing on the Centro (NUTS II) region over the period 2003–2022. Two architectures were implemented and tested: a multilayer perceptron (MLP) and a deep learning long short-term memory (DL-LSTM) model, trained and cross-validated on data from 2003–2021, with 2022 reserved as an independent test year.

Model skill was assessed both for routine conditions and during two well-documented extreme events: the 2020 Oleiros wildfires and the 2022 Serra da Estrela wildfires, which produced intense PM₁₀ episodes in central Portugal. The models showed high predictive capability for daily PM₁₀, with the MLP achieving a correlation coefficient of 0.97 and slightly outperforming the DL-LSTM configuration.

These results highlight the potential of data-driven methods to anticipate short-term air quality degradation, including wildfire-driven pollution peaks, and to support operational warning systems at the regional scale. The proposed framework can be extended to other pollutants and regions, contributing to more effective environmental management and public health planning in Portugal.