Design of an early alert system for PM2.5 through a stochastic method and machine learning models

Abstract

In Latin America, the levels of pollution have risen considerably in the last few years. 2019, for example, had one of the largest numbers of air quality alerts. These alerts signal an increase in respiratory diseases among the population. For this reason, this paper designs a preventive early alert system for air quality. This system compares three machine learning models and validates, through statistical and categorical parameters (9), that a stochastic model, combined with a convolution bidirectional recurrent neural network (1D-BDLM), has an accuracy of ≈93±4% when forecasting the risk for each population group in all the monitoring stations. Likewise, it is also able to capture high pollution events without producing false alarms (≈10 ± 5%). This model is utilized to design an alert protocol (24 h in advance) before a pollution event occurs. The protocol distinguishes the level of alert and the type of population at risk, focusing on two objectives: pollution mitigation and risk reduction for the population. To reduce pollutant concentrations, this paper proposes limiting vehicle traffic in the most polluted city zones or, if necessary, throughout the entire area. In relation to stationary sources, this article proposes the implementation of monitoring measures in order to identify the most polluting factories and restrict their operation during a specific period of time. In regards to population risk, the protocol aims to reduce exposure time by recommending the avoidance of outdoor activities (in specific zones) and the use of protective gear, taking into consideration relevant differences between population groups.