Prediction of PM2.5 and PM10 Concentrations Using XGBoost and LightGBM Algorithms: A Case Study in Lima, Peru

Air pollution is a major problem that affects both human health and the environment, causing millions of premature deaths annually worldwide and severely degrading the state of the planet. Exposure to fine particulate matter, which is highly hazardous, enables these particles to penetrate deeply into the lungs and lead to serious health issues, including a reduction in life expectancy by more than two years. In response to this problem, it is crucial to identify effective ways to monitor the levels of these pollutants in our daily surroundings. This article presents a case study conducted in the district of San Borja, Lima, Peru, where prediction models for PM2.5 and PM10 were implemented using the XGBoost and LightGBM algorithms. Employing data from the SENAMHI portal and a correlation analysis of variables, two different scenarios were developed for training the models. In scenario 1, prediction models for PM2.5 and PM10 were trained using all available meteorological and pollution variables. In scenario 2, the models were trained for PM2.5 excluding the PM10 variable, and vice versa. The results showed that both models achieved high accuracy, measured by the coefficient of determination, with no statistically significant difference indicating the superiority of either model. Furthermore, the analysis of the proposed scenarios revealed that excluding key variables can result in significantly less accurate predictions, potentially undermining the effectiveness of environmental management strategies.