A Methodology to Increase the Accuracy of Particulate Matter Predictors Based on Time Decomposition

Abstract

Particulate matter (PM) is one of the most harmful air pollutants to human health studied worldwide. In this scenario, it is of paramount importance to monitor and predict PM concentration. Artificial neural networks (ANN) are commonly used to forecast air pollution levels due to their accuracy. The use of partition on prediction problems is well known because decomposition of time series allows the latent components of the original series to be revealed. It is a matter of extracting the “deterministic” component, which is easy to predict the random components. However, there is no evidence of its use in air pollution forecasting. In this work, we introduce a different approach consisting of the decomposition of the time series in contiguous monthly partitions, aiming to develop specialized predictors to solve the problem because air pollutant concentration has seasonal behavior. The goal is to reach prediction accuracy higher than those obtained by using the entire series. Experiments were performed for seven time series of daily particulate matter concentrations (PM2.5 and PM10–particles with diameter less than 2.5 and 10 micrometers, respectively) in Finland and Brazil, using four ANNs: multilayer perceptron, radial basis function, extreme learning machines, and echo state networks. The experimental results using three evaluation measures showed that the proposed methodology increased all models’ prediction capability, leading to higher accuracy compared to the traditional approach, even for extremely high air pollution events. Our study has an important contribution to air quality prediction studies. It can help governments take measures aiming air pollution reduction and preparing hospitals during extreme air pollution events, which is related to the following United Nations sustainable developments goals: SDG 3—good health and well-being and SDG 11—sustainable cities and communities.