Artificial Intelligence and Multivariate Statistics in the Analysis of PM2.5 Behavior in the City of Barranquilla

Abstract

Predicting the behavior of atmospheric pollutants has become a crucial input for the design of government contingency strategies aiming to mitigate pollution levels. For this reason, this research presents a methodology to predict the non-compliance with the maximum allowed indices for the particulate matter pollutant PM2.5. The approach involves multivariate statistical models that can be used to understand how various meteorological variables influence the pollutant’s behavior. It also serves as a means of selecting variables for training a Bayesian neural network classifier. In addition to the multivariate model, a multifactorial analysis of variance was applied to assess the effect of day, hour, and month on the mean PM2.5 levels. The main results revealed that variables such as wind direction, temperature, humidity, atmospheric pressure, precipitation, solar radiation, time of day, day of the week, and month were statistically significant. These variables allowed the training of a neural network with an overall predictive capacity of 78.2%, reaching 81% accuracy in cases of noncompliance. This demonstrates that with manageable variables (meteorological variables, time, day, and month), a highly useful forecasting tool can be designed to contribute to air quality management.