Study and comparison of machine learning models for air PM 2.5 concentration prediction

Abstract

In the last several decades and as a result of various kinds of man-made activities, industrialization and human urbanization, the atmospheric environment pollution became a real threat to the human's health. The particles with a diameter of less than 2.5µm, one of the most harmful pollutants present in the air as it causes diseases in the respiratory system as well as cardiovascular ones. Consequently, it is beneficial to predict the particulate matter PM2.5 concentrations with high accuracy for the purpose of to alert people to make the right decision in order to fix the situation and improve the air quality especially in environments where it is essential. The prediction of the PM2.5 concentration have to pass throw a pre-processing stage then fed to the multiple models by passing a data chunk of twelve days to get the prediction for the next day. In this article, a comparative study between different Artificial Intelligence predictions models is presented: Bidirectional Long Short-Term Memory (Bi-LSTM), Time Distributed Convolutional Neural Network (CNN), and a hybrid model combining both CNN and Bi-LSTM. For this purpose, several architectures were used for the different models: Multi Inputs - Multi Outputs, Multi Inputs - Single Output and the univariate. The CNN extracts the internal spatial correlation between variables and the Bi-LSTM extracts the temporal patterns, the hybridization process proposed of those two models with the multiple Inputs -Single Output architecture gave us the most accurate results.