Comparative analysis of artificial intelligence algorithms for fine particulate matter prediction

Abstract This study aims to analyze the predictive performance of artificial intelligence algorithms in forecasting fine particulate matter (PM2.5) under different scenarios of temporal variability. PM2.5 data were collected over a year using low-cost sensors in ten sites, under urban and agricultural land use in central Argentina. Additionally, 178 time series of satellite variables were downloaded from the cloud to be used as predictors. Various machine learning models, including Linear Regression, Random Forest, and XGBoost, as well as deep learning models such as Long Short-Term Memory (LSTM), Gated Recurrent Units (GRUs) and Recurrent Neural Network (RNN), were compared. The hourly concentration of PM2.5 ranged from 0.53 \(\:{μ\:g}{{m}³}\) to 95.28 \(\:{μ\:g}{{m}³}\), with an average of 13.1 \(\:{μ\:g}{{m}³}\) for the urban land use and 6.78 \(\:{μ\:g}{{m}³}\) in the agricultural one. Temporal variability was highly different between scenarios, the interquantilic ranges were 7.31 and 4.38 \(\:{μ\:g}{{m}³}\) for the urban and agricultural land use, respectively. The GRU was the best algorithm in the urban land use (RMSE = 3.23 \(\:{μ\:g}{{m}³}\)). Conversely, for agricultural land use, the RF algorithm exhibited superior performance (RMSE = 2.54 \(\:{μ\:g}{{m}³}\)). The study findings highlight the strong impact of the time series variability and the learning capabilities of each predictive algorithm on predictions. Deep learning algorithms outperformed machine learning models in predicting PM2.5 from highly variable time series. Artificial intelligence algorithms demonstrate the capability for accurate forecasting of PM2.5 concentrations in the context of a low-cost alert system.